Introducing StylEx: A New Approach for Visual Explanation of Classifiers

Neural networks can perform certain tasks remarkably well, but understanding how they reach their decisions — e.g., identifying which signals in an image cause a model to determine it to be of one class and not another — is often a mystery. Explaining a neural model’s decision process may have high social impact in certain areas, such as analysis of medical images and autonomous driving, where human oversight is critical. These insights can also be helpful in guiding health care providers, revealing model biases, providing support for downstream decision makers, and even aiding scientific discovery.

Previous approaches for visual explanations of classifiers, such as attention maps (e.g., Grad-CAM), highlight which regions in an image affect the classification, but they do not explain what attributes within those regions determine the classification outcome: For example, is it their color? Their shape? Another family of methods provides an explanation by smoothly transforming the image between one class and another (e.g., GANalyze). However, these methods tend to change all attributes at once, thus making it difficult to isolate the individual affecting attributes.

In “Explaining in Style: Training a GAN to explain a classifier in StyleSpace”, presented at ICCV 2021, we propose a new approach for a visual explanation of classifiers. Our approach, StylEx, automatically discovers and visualizes disentangled attributes that affect a classifier. It allows exploring the effect of individual attributes by manipulating those attributes separately (changing one attribute does not affect others). StylEx is applicable to a wide range of domains, including animals, leaves, faces, and retinal images. Our results show that StylEx finds attributes that align well with semantic ones, generate meaningful image-specific explanations, and are interpretable by people as measured in user studies.

Explaining a Cat vs. Dog Classifier: StylEx provides the top-K discovered disentangled attributes which explain the classification. Moving each knob manipulates only the corresponding attribute in the image, keeping other attributes of the subject fixed.

For instance, to understand a cat vs. dog classifier on a given image, StylEx can automatically detect disentangled attributes and visualize how manipulating each attribute can affect the classifier probability. The user can then view these attributes and make semantic interpretations for what they represent. For example, in the figure above, one can draw conclusions such as “dogs are more likely to have their mouth open than cats” (attribute #4 in the GIF above), “cats’ pupils are more slit-like” (attribute #5), “cats’ ears do not tend to be folded” (attribute #1), and so on.

The video below provides a short explanation of the method:

How StylEx Works: Training StyleGAN to Explain a Classifier
Given a classifier and an input image, we want to find and visualize the individual attributes that affect its classification. For that, we utilize the StyleGAN2 architecture, which is known to generate high quality images. Our method consists of two phases:

Phase 1: Training StylEx

A recent work showed that StyleGAN2 contains a disentangled latent space called “StyleSpace”, which contains individual semantically meaningful attributes of the images in the training dataset. However, because StyleGAN training is not dependent on the classifier, it may not represent those attributes that are important for the decision of the specific classifier we want to explain. Therefore, we train a StyleGAN-like generator to satisfy the classifier, thus encouraging its StyleSpace to accommodate classifier-specific attributes.

This is achieved by training the StyleGAN generator with two additional components. The first is an encoder, trained together with the GAN with a reconstruction-loss, which forces the generated output image to be visually similar to the input. This allows us to apply the generator on any given input image. However, visual similarity of the image is not enough, as it may not necessarily capture subtle visual details important for a particular classifier (such as medical pathologies). To ensure this, we add a classification-loss to the StyleGAN training, which forces the classifier probability of the generated image to be the same as the classifier probability of the input image. This guarantees that subtle visual details important for the classifier (such as medical pathologies) will be included in the generated image.

Training StyleEx: We jointly train the generator and the encoder. A reconstruction-loss is applied between the generated image and the original image to preserve visual similarity. A classification-loss is applied between the classifier output of the generated image and the classifier output of the original image to ensure the generator captures subtle visual details important for the classification.

Phase 2: Extracting Disentangled Attributes

Once trained, we search the StyleSpace of the trained Generator for attributes that significantly affect the classifier. To do so, we manipulate each StyleSpace coordinate and measure its effect on the classification probability. We seek the top attributes that maximize the change in classification probability for the given image. This provides the top-K image-specific attributes. By repeating this process for a large number of images per class, we can further discover the top-K class-specific attributes, which teaches us what the classifier has learned about the specific class. We call our end-to-end system “StylEx”.

A visual illustration of image-specific attribute extraction: once trained, we search for the StyleSpace coordinates that have the highest effect on the classification probability of a given image.

StylEx is Applicable to a Wide Range of Domains and Classifiers
Our method works on a wide variety of domains and classifiers (binary and multi-class). Below are some examples of class-specific explanations. In all the domains tested, the top attributes detected by our method correspond to coherent semantic notions when interpreted by humans, as verified by human evaluation.

For perceived gender and age classifiers, below are the top four detected attributes per classifier. Our method exemplifies each attribute on multiple images that are automatically selected to best demonstrate that attribute. For each attribute we flicker between the source and attribute-manipulated image. The degree to which manipulating the attribute affects the classifier probability is shown at the top-left corner of each image.

Top-4 automatically detected attributes for a perceived-gender classifier.
Top-4 automatically detected attributes for a perceived-age classifier.

Note that our method explains a classifier, not reality. That is, the method is designed to reveal image attributes that a given classifier has learned to utilize from data; those attributes may not necessarily characterize actual physical differences between class labels (e.g., a younger or older age) in reality. In particular, these detected attributes may reveal biases in the classifier training or dataset, which is another key benefit of our method. It can further be used to improve fairness of neural networks, for example, by augmenting the training dataset with examples that compensate for the biases our method reveals.

Adding the classifier loss into StyleGAN training turns out to be crucial in domains where the classification depends on fine details. For example, a GAN trained on retinal images without a classifier loss will not necessarily generate fine pathological details corresponding to a particular disease. Adding the classification loss causes the GAN to generate these subtle pathologies as an explanation of the classifier. This is exemplified below for a retinal image classifier (DME disease) and a sick/healthy leaf classifier. StylEx is able to discover attributes that are aligned with disease indicators, for instance “hard exudates”, which is a well known marker for retinal DME, and rot for leaf diseases.

Top-4 automatically detected attributes for a DME classifier of retina images.
Top-4 automatically detected attributes for a classifier of sick/healthy leaf images.

Finally, this method is also applicable to multi-class problems, as demonstrated on a 200-way bird species classifier.

Top-4 automatically detected attributes in a 200-way classifier trained on CUB-2011 for (a) the class “brewer blackbird, and (b) the class yellow bellied flycatcher. Indeed we observe that StylEx detects attributes that correspond to attributes in CUB taxonomy.

Broader Impact and Next Steps
Overall, we have introduced a new technique that enables the generation of meaningful explanations for a given classifier on a given image or class. We believe that our technique is a promising step towards detection and mitigation of previously unknown biases in classifiers and/or datasets, in line with Google’s AI Principles. Additionally, our focus on multiple-attribute based explanation is key to providing new insights about previously opaque classification processes and aiding in the process of scientific discovery. Finally, our GitHub repository includes a Colab and model weights for the GANs used in our paper.

The research described in this post was done by Oran Lang, Yossi Gandelsman, Michal Yarom, Yoav Wald (as an intern), Gal Elidan, Avinatan Hassidim, William T. Freeman, Phillip Isola, Amir Globerson, Michal Irani and Inbar Mosseri. We would like to thank Jenny Huang and Marilyn Zhang for leading the writing process for this blogpost, and Reena Jana, Paul Nicholas, and Johnny Soraker for ethics reviews of our research paper and this post.


Learning to Route by Task for Efficient Inference

Scaling large language models has resulted in significant quality improvements natural language understanding (T5), generation (GPT-3) and multilingual neural machine translation (M4). One common approach to building a larger model is to increase the depth (number of layers) and width (layer dimensionality), simply enlarging existing dimensions of the network. Such dense models take an input sequence (divided into smaller components, called tokens) and pass every token through the full network, activating every layer and parameter. While these large, dense models have achieved state-of-the-art results on multiple natural language processing (NLP) tasks, their training cost increases linearly with model size.

An alternative, and increasingly popular, approach is to build sparsely activated models based on a mixture of experts (MoE) (e.g., GShard-M4 or GLaM), where each token passed to the network follows a separate subnetwork by skipping some of the model parameters. The choice of how to distribute the input tokens to each subnetwork (the “experts”) is determined by small router networks that are trained together with the rest of the network. This allows researchers to increase model size (and hence, performance) without a proportional increase in training cost.

While this is an effective strategy at training time, sending tokens of a long sequence to multiple experts, again makes inference computationally expensive because the experts have to be distributed among a large number of accelerators. For example, serving the 1.2T parameter GLaM model requires 256 TPU-v3 chips. Much like dense models, the number of processors needed to serve an MoE model still scales linearly with respect to the model size, increasing compute requirements while also resulting in significant communication overhead and added engineering complexity.

In “Beyond Distillation: Task-level Mixture-of-Experts for Efficient Inference”, we introduce a method called Task-level Mixture-of-Experts (TaskMoE), that takes advantage of the quality gains of model scaling while still being efficient to serve. Our solution is to train a large multi-task model from which we then extract smaller, stand-alone per-task subnetworks suitable for inference with no loss in model quality and with significantly reduced inference latency. We demonstrate the effectiveness of this method for multilingual neural machine translation (NMT) compared to other mixture of experts models and to models compressed using knowledge distillation.

Training Large Sparsely Activated Models with Task Information
We train a sparsely activated model, where router networks learn to send tokens of each task-specific input to different subnetworks of the model associated with the task of interest. For example, in the case of multilingual NMT, every token of a given language is routed to the same subnetwork. This differs from other recent approaches, such as the sparsely gated mixture of expert models (e.g., TokenMoE), where router networks learn to send different tokens in an input to different subnetworks independent of task.

Inference: Bypassing Distillation by Extracting Subnetworks
A consequence of this difference in training between TaskMoE and models like TokenMoE is in how we approach inference. Because TokenMoE follows the practice of distributing tokens of the same task to many experts at both training and inference time, it is still computationally expensive at inference.

For TaskMoE, we dedicate a smaller subnetwork to a single task identity during training and inference. At inference time, we extract subnetworks by discarding unused experts for each task. TaskMoE and its variants enable us to train a single large multi-task network and then use a separate subnetwork at inference time for each task without using any additional compression methods post-training. We illustrate the process of training a TaskMoE network and then extracting per-task subnetworks for inference below.

During training, tokens of the same language are routed to the same expert based on language information (either source, target or both) in task-based MoE. Later, during inference we extract subnetworks for each task and discard unused experts.

To demonstrate this approach, we train models based on the Transformer architecture. Similar to GShard-M4 and GLaM, we replace the feedforward network of every other transformer layer with a Mixture-of-Experts (MoE) layer that consists of multiple identical feedforward networks, the “experts”. For each task, the routing network, trained along with the rest of the model, keeps track of the task identity for all input tokens and chooses a certain number of experts per layer (two in this case) to form the task-specific subnetwork. The baseline dense Transformer model has 143M parameters and 6 layers on both the encoder and decoder. The TaskMoE and TokenMoE that we train are also both 6 layers deep but with 32 experts for every MoE layer and have a total of 533M parameters. We train our models using publicly available WMT datasets, with over 431M sentences across 30 language pairs from different language families and scripts. We point the reader to the full paper for further details.

In order to demonstrate the advantage of using TaskMoE at inference time, we compare the throughput, or the number of tokens decoded per second, for TaskMoE, TokenMoE, and a baseline dense model. Once the subnetwork for each task is extracted, TaskMoE is 7x smaller than the 533M parameter TokenMoE model, and it can be served on a single TPUv3 core, instead of 64 cores required for TokenMoE. We see that TaskMoE has a peak throughput twice as high as that of TokenMoE models. In addition, on inspecting the TokenMoE model, we find that 25% of the inference time has been spent in inter-device communication, while virtually no time is spent in communication by TaskMoE.

Comparing the throughput of TaskMoE with TokenMoE across different batch sizes. The maximum batch size for TokenMoE is 1024 as opposed to 4096 for TaskMoE and the dense baseline model. Here, TokenMoE has one instance distributed across 64 TPUv3 cores, while TaskMoE and the baseline model have one instance on each of the 64 cores.

A popular approach to building a smaller network that still performs well is through knowledge distillation, in which a large teacher model trains a smaller student model with the goal of matching the teacher’s performance. However, this method comes at the cost of additional computation needed to train the student from the teacher. So, we also compare TaskMoE to a baseline TokenMoE model that we compress using knowledge distillation. The compressed TokenMoE model has a size comparable to the per-task subnetwork extracted from TaskMoE.

We find that in addition to being a simpler method that does not need any additional training, TaskMoE improves upon a distilled TokenMoE model by 2.1 BLEU on average across all languages in our multilingual translation model. We note that distillation retains 43% of the performance gains achieved from scaling a dense multilingual model to a TokenMoE, whereas extracting the smaller subnetwork from the TaskMoE model results in no loss of quality.

BLEU scores (higher is better) comparing a distilled TokenMoE model to the TaskMoE and TokenMoE models with 12 layers (6 on the encoder and 6 on the decoder) and 32 experts. While both approaches improve upon a multilingual dense baseline, TaskMoE improves upon the baseline by 3.1 BLEU on average while distilling from TokenMoE improves upon the baseline by 1.0 BLEU on average.

Next Steps
The quality improvements often seen with scaling machine learning models has incentivized the research community to work toward advancing scaling technology to enable efficient training of large models. The emerging need to train models capable of generalizing to multiple tasks and modalities only increases the need for scaling models even further. However, the practicality of serving these large models remains a major challenge. Efficiently deploying large models is an important direction of research, and we believe TaskMoE is a promising step towards more inference friendly algorithms that retain the quality gains of scaling.

We would like to first thank our coauthors – Yanping Huang, Ankur Bapna, Maxim Krikun, Dmitry Lepikhin and Minh-Thang Luong. We would also like to thank Wolfgang Macherey, Yuanzhong Xu, Zhifeng Chen and Macduff Richard Hughes for their helpful feedback. Special thanks to the Translate and Brain teams for their useful input and discussions, and the entire GShard development team for their foundational contributions to this project. We would also like to thank Tom Small for creating the animations for the blog post.


Scaling Vision with Sparse Mixture of Experts

Advances in deep learning over the last few decades have been driven by a few key elements. With a small number of simple but flexible mechanisms (i.e., inductive biases such as convolutions or sequence attention), increasingly large datasets, and more specialized hardware, neural networks can now achieve impressive results on a wide range of tasks, such as image classification, machine translation, and protein folding prediction.

However, the use of large models and datasets comes at the expense of significant computational requirements. Yet, recent works suggest that large model sizes might be necessary for strong generalization and robustness, so training large models while limiting resource requirements is becoming increasingly important. One promising approach involves the use of conditional computation: rather than activating the whole network for every single input, different parts of the model are activated for different inputs. This paradigm has been featured in the Pathways vision and recent works on large language models, while it has not been well explored in the context of computer vision.

In “Scaling Vision with Sparse Mixture of Experts”, we present V-MoE, a new vision architecture based on a sparse mixture of experts, which we then use to train the largest vision model to date. We transfer V-MoE to ImageNet and demonstrate matching state-of-the-art accuracy while using about 50% fewer resources than models of comparable performance. We have also open-sourced the code to train sparse models and provided several pre-trained models.

Vision Mixture of Experts (V-MoEs)
Vision Transformers (ViT) have emerged as one of the best architectures for vision tasks. ViT first partitions an image into equally-sized square patches. These are called tokens, a term inherited from language models. Still, compared to the largest language models, ViT models are several orders of magnitude smaller in terms of number of parameters and compute.

To massively scale vision models, we replace some dense feedforward layers (FFN) in the ViT architecture with a sparse mixture of independent FFNs (which we call experts). A learnable router layer selects which experts are chosen (and how they are weighted) for every individual token. That is, different tokens from the same image may be routed to different experts. Each token is only routed to at most K (typically 1 or 2) experts, among a total of E experts (in our experiments, E is typically 32). This allows scaling the model’s size while keeping its computation per token roughly constant. The figure below shows the structure of the encoder blocks in more detail.

V-MoE Transformer Encoder block.

Experimental Results
We first pre-train the model once on JFT-300M, a large dataset of images. The left plot below shows our pre-training results for models of all sizes: from the small S/32 to the huge H/14.

We then transfer the model to new downstream tasks (such as ImageNet), by using a new head (the last layer in a model). We explore two transfer setups: either fine-tuning the entire model on all available examples of the new task, or freezing the pre-trained network and tuning only the new head using a few examples (known as few-shot transfer). The right plot in the figure below summarizes our transfer results to ImageNet, training on only 5 images per class (called 5-shot transfer).

JFT-300M Precision@1 and ImageNet 5-shot accuracy. Colors represent different ViT variants and markers represent either standard ViT (●), or V-MoEs (▸) with expert layers on the last n even blocks. We set n=2 for all models, except V-MoE-H where n=5. Higher indicates better performance, with more efficient models being to the left.

In both cases, the sparse model strongly outperforms its dense counterpart at a given amount of training compute (shown by the V-MoE line being above the ViT line), or achieves similar performance much faster (shown by the V-MoE line being to the left of the ViT line).

To explore the limits of vision models, we trained a 15-billion parameter model with 24 MoE layers (out of 48 blocks) on an extended version of JFT-300M. This massive model — the largest to date in vision as far as we know — achieved 90.35% test accuracy on ImageNet after fine-tuning, near the current state-of-the-art.

Priority Routing
In practice, due to hardware constraints, it is not efficient to use buffers with a dynamic size, so models typically use a pre-defined buffer capacity for each expert. Assigned tokens beyond this capacity are dropped and not processed once the expert becomes “full”. As a consequence, higher capacities yield higher accuracy, but they are also more computationally expensive.

We leverage this implementation constraint to make V-MoEs faster at inference time. By decreasing the total combined buffer capacity below the number of tokens to be processed, the network is forced to skip processing some tokens in the expert layers. Instead of choosing the tokens to skip in some arbitrary fashion (as previous works did), the model learns to sort tokens according to an importance score. This maintains high quality predictions while saving a lot of compute. We refer to this approach as Batch Priority Routing (BPR), illustrated below.

Under high capacity, both vanilla and priority routing work well as all patches are processed. However, when the buffer size is reduced to save compute, vanilla routing selects arbitrary patches to process, often leading to poor predictions. BPR smartly prioritizes important patches resulting in better predictions at lower computational costs.

Dropping the right tokens turns out to be essential to deliver high-quality and more efficient inference predictions. When the expert capacity decreases, performance quickly decreases with the vanilla routing mechanism. Conversely, BPR is much more robust to low capacities.

Performance versus inference capacity buffer size (or ratio) C for a V-MoE-H/14 model with K=2. Even for large C’s, BPR improves performance; at low C the difference is quite significant. BPR is competitive with dense models (ViT-H/14) by processing only 15-30% of the tokens.

Overall, we observed that V-MoEs are highly flexible at inference time: for instance, one can decrease the number of selected experts per token to save time and compute, without any further training on the model weights.

Exploring V-MoEs
Because much is yet to be discovered about the internal workings of sparse networks, we also explored the routing patterns of the V-MoE.

One hypothesis is that routers would learn to discriminate and assign tokens to experts based on some semantic grounds (the “car” expert, the “animal” experts, and so on). To test this, below we show plots for two different MoE layers (a very early-on one, and another closer to the head). The x-axis corresponds to each of the 32 experts, and the y-axis shows the ID of the image classes (from 1 to 1000). Each entry in the plot shows how often an expert was selected for tokens corresponding to the specific image class, with darker colors indicating higher frequency. While in the early layers there is little correlation, later in the network, each expert receives and processes tokens from only a handful of classes. Therefore, we can conclude that some semantic clustering of the patches emerges in the deeper layers of the network.

Higher routing decisions correlate with image classes. We show two MoE layers of a V-MoE-H/14. The x-axis corresponds to the 32 experts in a layer. The y-axis are the 1000 ImageNet classes; orderings for both axes are different across plots (to highlight correlations). For each pair (expert e, class c) we show the average routing weight for the tokens corresponding to all images with class c for that particular expert e.

Final Thoughts
We train very large vision models using conditional computation, delivering significant improvements in representation and transfer learning for relatively little training cost. Alongside V-MoE, we introduced BPR, which requires the model to process only the most useful tokens in the expert layers.

We believe this is just the beginning of conditional computation at scale for computer vision; extensions include multi-modal and multi-task models, scaling up the expert count, and improving transfer of the representations produced by sparse models. Heterogeneous expert architectures and conditional variable-length routes are also promising directions. Sparse models can especially help in data rich domains such as large-scale video modeling. We hope our open-source code and models help attract and engage researchers new to this field.

We thank our co-authors: Basil Mustafa, Maxim Neumann, Rodolphe Jenatton, André Susano Pinto, Daniel Keysers, and Neil Houlsby. We thank Alex Kolesnikov, Lucas Beyer, and Xiaohua Zhai for providing continuous help and details about scaling ViT models. We are also grateful to Josip Djolonga, Ilya Tolstikhin, Liam Fedus, and Barret Zoph for feedback on the paper; James Bradbury, Roy Frostig, Blake Hechtman, Dmitry Lepikhin, Anselm Levskaya, and Parker Schuh for invaluable support helping us run our JAX models efficiently on TPUs; and many others from the Brain team for their support. Finally, we would also like to thank and acknowledge Tom Small for the awesome animated figure used in this post.


Google Research: Themes from 2021 and Beyond

Over the last several decades, I’ve witnessed a lot of change in the fields of machine learning (ML) and computer science. Early approaches, which often fell short, eventually gave rise to modern approaches that have been very successful. Following that long-arc pattern of progress, I think we’ll see a number of exciting advances over the next several years, advances that will ultimately benefit the lives of billions of people with greater impact than ever before. In this post, I’ll highlight five areas where ML is poised to have such impact. For each, I’ll discuss related research (mostly from 2021) and the directions and progress we’ll likely see in the next few years.



  · Trend 1: More Capable, General-Purpose ML Models
  · Trend 2: Continued Efficiency Improvements for ML
  · Trend 3: ML Is Becoming More Personally and Communally Beneficial
  · Trend 4: Growing Benefits of ML in Science, Health and Sustainability
  · Trend 5: Deeper and Broader Understanding of ML

Trend 1: More Capable, General-Purpose ML Models
Researchers are training larger, more capable machine learning models than ever before. For example, just in the last couple of years models in the language domain have grown from billions of parameters trained on tens of billions of tokens of data (e.g., the 11B parameter T5 model), to hundreds of billions or trillions of parameters trained on trillions of tokens of data (e.g., dense models such as OpenAI’s 175B parameter GPT-3 model and DeepMind’s 280B parameter Gopher model, and sparse models such as Google’s 600B parameter GShard model and 1.2T parameter GLaM model). These increases in dataset and model size have led to significant increases in accuracy for a wide variety of language tasks, as shown by across-the-board improvements on standard natural language processing (NLP) benchmark tasks (as predicted by work on neural scaling laws for language models and machine translation models).

Many of these advanced models are focused on the single but important modality of written language and have shown state-of-the-art results in language understanding benchmarks and open-ended conversational abilities, even across multiple tasks in a domain. They have also shown exciting capabilities to generalize to new language tasks with relatively little training data, in some cases, with few to no training examples for a new task. A couple of examples include improved long-form question answering, zero-label learning in NLP, and our LaMDA model, which demonstrates a sophisticated ability to carry on open-ended conversations that maintain significant context across multiple turns of dialog.

A dialog with LaMDA mimicking a Weddell seal with the preset grounding prompt, “Hi I’m a weddell seal. Do you have any questions for me?” The model largely holds down a dialog in character.
(Weddell Seal image cropped from Wikimedia CC licensed image.)

Transformer models are also having a major impact in image, video, and speech models, all of which also benefit significantly from scale, as predicted by work on scaling laws for visual transformer models. Transformers for image recognition and for video classification are achieving state-of-the-art results on many benchmarks, and we’ve also demonstrated that co-training models on both image data and video data can improve performance on video tasks compared with video data alone. We’ve developed sparse, axial attention mechanisms for image and video transformers that use computation more efficiently, found better ways of tokenizing images for visual transformer models, and improved our understanding of visual transformer methods by examining how they operate compared with convolutional neural networks. Combining transformer models with convolutional operations has shown significant benefits in visual as well as speech recognition tasks.

The outputs of generative models are also substantially improving. This is most apparent in generative models for images, which have made significant strides over the last few years. For example, recent models have demonstrated the ability to create realistic images given just a category (e.g., “irish setter” or “streetcar”, if you desire), can “fill in” a low-resolution image to create a natural-looking high-resolution counterpart (“computer, enhance!”), and can even create natural-looking aerial nature scenes of arbitrary length. As another example, images can be converted to a sequence of discrete tokens that can then be synthesized at high fidelity with an autoregressive generative model.

Example of a cascade diffusion models that generate novel images from a given category and then use those as the seed to create high-resolution examples: the first model generates a low resolution image, and the rest perform upsampling to the final high resolution image.
The SR3 super-resolution diffusion model takes as input a low-resolution image, and builds a corresponding high resolution image from pure noise.

Because these are powerful capabilities that come with great responsibility, we carefully vet potential applications of these sorts of models against our AI Principles.

Beyond advanced single-modality models, we are also starting to see large-scale multi-modal models. These are some of the most advanced models to date because they can accept multiple different input modalities (e.g., language, images, speech, video) and, in some cases, produce different output modalities, for example, generating images from descriptive sentences or paragraphs, or describing the visual content of images in human languages. This is an exciting direction because like the real world, some things are easier to learn in data that is multimodal (e.g., reading about something and seeing a demonstration is more useful than just reading about it). As such, pairing images and text can help with multi-lingual retrieval tasks, and better understanding of how to pair text and image inputs can yield improved results for image captioning tasks. Similarly, jointly training on visual and textual data can also help improve accuracy and robustness on visual classification tasks, while co-training on image, video, and audio tasks improves generalization performance for all modalities. There are also tantalizing hints that natural language can be used as an input for image manipulation, telling robots how to interact with the world and controlling other software systems, portending potential changes to how user interfaces are developed. Modalities handled by these models will include speech, sounds, images, video, and languages, and may even extend to structured data, knowledge graphs, and time series data.

Example of a vision-based robotic manipulation system that is able to generalize to novel tasks. Left: The robot is performing a task described in natural language to the robot as “place grapes in ceramic bowl”, without the model being trained on that specific task. Right: As on the left, but with the novel task description of “place bottle in tray”.

Often these models are trained using self-supervised learning approaches, where the model learns from observations of “raw” data that has not been curated or labeled, e.g., language models used in GPT-3 and GLaM, the self-supervised speech model BigSSL, the visual contrastive learning model SimCLR, and the multimodal contrastive model VATT. Self-supervised learning allows a large speech recognition model to match the previous Voice Search automatic speech recognition (ASR) benchmark accuracy while using only 3% of the annotated training data. These trends are exciting because they can substantially reduce the effort required to enable ML for a particular task, and because they make it easier (though by no means trivial) to train models on more representative data that better reflects different subpopulations, regions, languages, or other important dimensions of representation.

All of these trends are pointing in the direction of training highly capable general-purpose models that can handle multiple modalities of data and solve thousands or millions of tasks. By building in sparsity, so that the only parts of a model that are activated for a given task are those that have been optimized for it, these multimodal models can be made highly efficient. Over the next few years, we are pursuing this vision in a next-generation architecture and umbrella effort called Pathways. We expect to see substantial progress in this area, as we combine together many ideas that to date have been pursued relatively independently.

Pathways: a depiction of a single model we are working towards that can generalize across millions of tasks.


Trend 2: Continued Efficiency Improvements for ML
Improvements in efficiency — arising from advances in computer hardware design as well as ML algorithms and meta-learning research — are driving greater capabilities in ML models. Many aspects of the ML pipeline, from the hardware on which a model is trained and executed to individual components of the ML architecture, can be optimized for efficiency while maintaining or improving on state-of-the-art performance overall. Each of these different threads can improve efficiency by a significant multiplicative factor, and taken together, can reduce computational costs, including CO2 equivalent emissions (CO2e), by orders of magnitude compared to just a few years ago. This greater efficiency has enabled a number of critical advances that will continue to dramatically improve the efficiency of machine learning, enabling larger, higher quality ML models to be developed cost effectively and further democratizing access. I’m very excited about these dirctions of research!

Continued Improvements in ML Accelerator Performance

Each generation of ML accelerator improves on previous generations, enabling faster performance per chip, and often increasing the scale of the overall systems. Last year, we announced our TPUv4 systems, the fourth generation of Google’s Tensor Processing Unit, which demonstrated a 2.7x improvement over comparable TPUv3 results in the MLPerf benchmarks. Each TPUv4 chip has ~2x the peak performance per chip versus the TPUv3 chip, and the scale of each TPUv4 pod is 4096 chips (4x that of TPUv3 pods), yielding a performance of approximately 1.1 exaflops per pod (versus ~100 petaflops per TPUv3 pod). Having pods with larger numbers of chips that are connected together with high speed networks improves efficiency for larger models.

ML capabilities on mobile devices are also increasing significantly. The Pixel 6 phone features a brand new Google Tensor processor that integrates a powerful ML accelerator to better support important on-device features.

Left: TPUv4 board; Center: Part of a TPUv4 pod; Right: Google Tensor chip found in Pixel 6 phones.

Our use of ML to accelerate the design of computer chips of all kinds (more on this below) is also paying dividends, particularly to produce better ML accelerators.

Continued Improvements in ML Compilation and Optimization of ML Workloads

Even when the hardware is unchanged, improvements in compilers and other optimizations in system software for machine learning accelerators can lead to significant improvements in efficiency. For example, “A Flexible Approach to Autotuning Multi-pass Machine Learning Compilers” shows how to use machine learning to perform auto-tuning of compilation settings to get across-the-board performance improvements of 5-15% (and sometimes as much as 2.4x improvement) for a suite of ML programs on the same underlying hardware. GSPMD describes an automatic parallelization system based on the XLA compiler that is capable of scaling most deep learning network architectures beyond the memory capacity of an accelerator and has been applied to many large models, such as GShard-M4, LaMDA, BigSSL, ViT, MetNet-2, and GLaM, leading to state-of-the-art results across several domains.

End-to-end model speedups from using ML-based compiler autotuning on 150 ML models. Included are models that achieve improvements of 5% or more. Bar colors represent relative improvement from optimizing different model components.

Human-Creativity–Driven Discovery of More Efficient Model Architectures

Continued improvements in model architectures give substantial reductions in the amount of computation needed to achieve a given level of accuracy for many problems. For example, the Transformer architecture, which we developed in 2017, was able to improve the state of the art on several NLP and translation benchmarks while simultaneously using 10x to 100x less computation to achieve these results than a variety of other prevalent methods, such as LSTMs and other recurrent architectures. Similarly, the Vision Transformer was able to show improved state-of-the-art results on a number of different image classification tasks despite using 4x to 10x less computation than convolutional neural networks.

Machine-Driven Discovery of More Efficient Model Architectures

Neural architecture search (NAS) can automatically discover new ML architectures that are more efficient for a given problem domain. A primary advantage of NAS is that it can greatly reduce the effort needed for algorithm development, because NAS requires only a one-time effort per search space and problem domain combination. In addition, while the initial effort to perform NAS can be computationally expensive, the resulting models can greatly reduce computation in downstream research and production settings, resulting in greatly reduced resource requirements overall. For example, the one-time search to discover the Evolved Transformer generated only 3.2 tons of CO2e (much less than the 284t CO2e reported elsewhere; see Appendix C and D in this joint Google/UC Berkeley preprint), but yielded a model for use by anyone in the NLP community that is 15-20% more efficient than the plain Transformer model. A more recent use of NAS discovered an even more efficient architecture called Primer (that has also been open-sourced), which reduces training costs by 4x compared to a plain Transformer model. In this way, the discovery costs of NAS searches are often recouped from the use of the more-efficient model architectures that are discovered, even if they are applied to only a handful of downstream uses (and many NAS results are reused thousands of times).

The Primer architecture discovered by NAS is 4x as efficient compared with a plain Transformer model. This image shows (in red) the two main modifications that give Primer most of its gains: depthwise convolution added to attention multi-head projections and squared ReLU activations (blue indicates portions of the original Transformer).

NAS has also been used to discover more efficient models in the vision domain. The EfficientNetV2 model architecture is the result of a neural architecture search that jointly optimizes for model accuracy, model size, and training speed. On the ImageNet benchmark, EfficientNetV2 improves training speed by 5–11x while substantially reducing model size over previous state-of-the-art models. The CoAtNet model architecture was created with an architecture search that uses ideas from the Vision Transformer and convolutional networks to create a hybrid model architecture that trains 4x faster than the Vision Transformer and achieves a new ImageNet state of the art.

EfficientNetV2 achieves much better training efficiency than prior models for ImageNet classification.

The broad use of search to help improve ML model architectures and algorithms, including the use of reinforcement learning and evolutionary techniques, has inspired other researchers to apply this approach to different domains. To aid others in creating their own model searches, we have open-sourced Model Search, a platform that enables others to explore model search for their domains of interest. In addition to model architectures, automated search can also be used to find new, more efficient reinforcement learning algorithms, building on the earlier AutoML-Zero work that demonstrated this approach for automating supervised learning algorithm discovery.

Use of Sparsity

Sparsity, where a model has a very large capacity, but only some parts of the model are activated for a given task, example or token, is another important algorithmic advance that can greatly improve efficiency. In 2017, we introduced the sparsely-gated mixture-of-experts layer, which demonstrated better results on a variety of translation benchmarks while using 10x less computation than previous state-of-the-art dense LSTM models. More recently, Switch Transformers, which pair a mixture-of-experts–style architecture with the Transformer model architecture, demonstrated a 7x speedup in training time and efficiency over the dense T5-Base Transformer model. The GLaM model showed that transformers and mixture-of-expert–style layers can be combined to produce a model that exceeds the accuracy of the GPT-3 model on average across 29 benchmarks using 3x less energy for training and 2x less computation for inference. The notion of sparsity can also be applied to reduce the cost of the attention mechanism in the core Transformer architecture.

The BigBird sparse attention model consists of global tokens that attend to all parts of an input sequence, local tokens, and a set of random tokens. Theoretically, this can be interpreted as adding a few global tokens on a Watts-Strogatz graph.

The use of sparsity in models is clearly an approach with very high potential payoff in terms of computational efficiency, and we are only scratching the surface in terms of research ideas to be tried in this direction.

Each of these approaches for improved efficiency can be combined together so that equivalent-accuracy language models trained today in efficient data centers are ~100 times more energy efficient and produce ~650 times less CO2e emissions, compared to a baseline Transformer model trained using P100 GPUs in an average U.S. datacenter using an average U.S. energy mix. And this doesn’t even account for Google’s carbon-neutral, 100% renewable energy offsets. We’ll have a more detailed blog post analyzing the carbon emissions trends of NLP models soon.


Trend 3: ML Is Becoming More Personally and Communally Beneficial
A host of new experiences are made possible as innovation in ML and silicon hardware (like the Google Tensor processor on the Pixel 6) enable mobile devices to be more capable of continuously and efficiently sensing their surrounding context and environment. These advances have improved accessibility and ease of use, while also boosting computational power, which is critical for popular features like mobile photography, live translation and more. Remarkably, recent technological advances also provide users with a more customized experience while strengthening privacy safeguards.

More people than ever rely on their phone cameras to record their daily lives and for artistic expression. The clever application of ML to computational photography has continued to advance the capabilities of phone cameras, making them easier to use, improving performance, and resulting in higher-quality images. Advances, such as improved HDR+, the ability to take pictures in very low light, better handling of portraits, and efforts to make cameras more inclusive so they work for all skin tones, yield better photos that are more true to the photographer’s vision and to their subjects. Such photos can be further improved using the powerful ML-based tools now available in Google Photos, like cinematic photos, noise and blur reduction, and the Magic Eraser.

HDR+ starts from a burst of full-resolution raw images, each underexposed by the same amount (left). The merged image has reduced noise and increased dynamic range, leading to a higher quality final result (right).

In addition to using their phones for creative expression, many people rely on them to help communicate with others across languages and modalities in real-time using Live Translate in messaging apps and Live Caption for phone calls. Speech recognition accuracy has continued to make substantial improvements thanks to techniques like self-supervised learning and noisy student training, with marked improvements for accented speech, noisy conditions or environments with overlapping speech, and across many languages. Building on advances in text-to-speech synthesis, people can listen to web pages and articles using our Read Aloud technology on a growing number of platforms, making information more available across barriers of modality and languages. Live speech translations in the Google Translate app have become significantly better by stabilizing the translations that are generated on-the-fly, and high quality, robust and responsible direct speech-to-speech translation provides a much better user experience in communicating with people speaking a different language. New work on combining ML with traditional codec approaches in the Lyra speech codec and the more general SoundStream audio codec enables higher fidelity speech, music, and other sounds to be communicated reliably at much lower bitrate.

Everyday interactions are becoming much more natural with features like automatic call screening and ML agents that will wait on hold for you, thanks to advances in Duplex. Even short tasks that users may perform frequently have been improved with tools such as Smart Text Selection, which automatically selects entities like phone numbers or addresses for easy copy and pasting, and grammar correction as you type on Pixel 6 phones. In addition, Screen Attention prevents the phone screen from dimming when you are looking at it and improvements in gaze recognition are opening up new use cases for accessibility and for improved wellness and health. ML is also enabling new methods for ensuring the safety of people and communities. For example, Suspicious Message Alerts warn against possible phishing attacks and Safer Routing detects hard-braking events to suggest alternate routes.

Recent work demonstrates the ability of gaze recognition as an important biomarker of mental fatigue.

Given the potentially sensitive nature of the data that underlies these new capabilities, it is essential that they are designed to be private by default. Many of them run inside of Android’s Private Compute Core — an open source, secure environment isolated from the rest of the operating system. Android ensures that data processed in the Private Compute Core is not shared to any apps without the user taking an action. Android also prevents any feature inside the Private Compute Core from having direct access to the network. Instead, features communicate over a small set of open-source APIs to Private Compute Services, which strips out identifying information and makes use of privacy technologies, including federated learning, federated analytics, and private information retrieval, enabling learning while simultaneously ensuring privacy.

Federated Reconstruction is a novel partially local federated learning technique in which models are partitioned into global and local parameters. For each round of Federated Reconstruction training: (1) The server sends the current global parameters g to each user i; (2) Each user i freezes g and reconstructs their local parameters li; (3) Each user i freezes li and updates g to produce gi; (4) Users’ gi are averaged to produce the global parameters for the next round.

These technologies are critical to evolving next-generation computation and interaction paradigms, whereby personal or communal devices can both learn from and contribute to training a collective model of the world without compromising privacy. A federated unsupervised approach to privately learn the kinds of aforementioned general-purpose models with fine-tuning for a given task or context could unlock increasingly intelligent systems that are far more intuitive to interact with — more like a social entity than a machine. Broad and equitable access to these intelligent interfaces will only be possible with deep changes to our technology stacks, from the edge to the datacenter, so that they properly support neural computing.


Trend 4: Growing Impact of ML in Science, Health and Sustainability
In recent years, we have seen an increasing impact of ML in the basic sciences, from physics to biology, with a number of exciting practical applications in related realms, such as renewable energy and medicine. Computer vision models have been deployed to address problems at both personal and global scales. They can assist physicians in their regular work, expand our understanding of neural physiology, and also provide better weather forecasts and streamline disaster relief efforts. Other types of ML models are proving critical in addressing climate change by discovering ways to reduce emissions and improving the output of alternative energy sources. Such models can even be leveraged as creative tools for artists! As ML becomes more robust, well-developed, and widely accessible, its potential for high-impact applications in a broad array of real-world domains continues to expand, helping to solve some of our most challenging problems.

Large-Scale Application of Computer Vision for New Insights

The advances in computer vision over the past decade have enabled computers to be used for a wide variety of tasks across different scientific domains. In neuroscience, automated reconstruction techniques can recover the neural connective structure of brain tissues from high resolution electron microscopy images of thin slices of brain tissue. In previous years, we have collaborated to create such resources for fruit fly, mouse, and songbird brains, but last year, we collaborated with the Lichtman Lab at Harvard University to analyze the largest sample of brain tissue imaged and reconstructed in this level of detail, in any species, and produced the first large-scale study of synaptic connectivity in the human cortex that spans multiple cell types across all layers of the cortex. The goal of this work is to produce a novel resource to assist neuroscientists in studying the stunning complexity of the human brain. The image below, for example, shows six neurons out of about 86 billion neurons in an adult human brain.

A single human chandelier neuron from our human cortex reconstruction, along with some of the pyramidal neurons that make a connection with that cell. Here’s an interactive version and a gallery of other interactive examples.

Computer vision technology also provides powerful tools to address challenges at much larger, even global, scales. A deep-learning–based approach to weather forecasting that uses satellite and radar imagery as inputs, combined with other atmospheric data, produces weather and precipitation forecasts that are more accurate than traditional physics-based models at forecasting times up to 12 hours. They can also produce updated forecasts much more quickly than traditional methods, which can be critical in times of extreme weather.

Comparison of 0.2 mm/hr precipitation on March 30, 2020 over Denver, Colorado. Left: Ground truth, source MRMS. Center: Probability map as predicted by MetNet-2. Right: Probability map as predicted by the physics-based HREF model. MetNet-2 is able to predict the onset of the storm earlier in the forecast than HREF as well as the storm’s starting location, whereas HREF misses the initiation location, but captures its growth phase well.

Having an accurate record of building footprints is essential for a range of applications, from population estimation and urban planning to humanitarian response and environmental science. In many parts of the world, including much of Africa, this information wasn’t previously available, but new work shows that using computer vision techniques applied to satellite imagery can help identify building boundaries at continental scales. The results of this approach have been released in the Open Buildings dataset, a new open-access data resource that contains the locations and footprints of 516 million buildings with coverage across most of the African continent. We’ve also been able to use this unique dataset in our collaboration with the World Food Programme to provide fast damage assessment after natural disasters through application of ML.

Example of segmenting buildings in satellite imagery. Left: Source image; Center: Semantic segmentation, with each pixel assigned a confidence score that it is a building vs. non-building; Right: Instance segmentation, obtained by thresholding and grouping together connected components.

A common theme across each of these cases is that ML models are able to perform specialized tasks efficiently and accurately based on analysis of available visual data, supporting high impact downstream tasks.

Automated Design Space Exploration

Another approach that has yielded excellent results across many fields is to allow an ML algorithm to explore and evaluate a problem’s design space for possible solutions in an automated way. In one application, a Transformer-based variational autoencoder learns to create aesthetically-pleasing and useful document layouts, and the same approach can be extended to explore possible furniture layouts. Another ML-driven approach automates the exploration of the huge design space of tweaks for computer game rules to improve playability and other attributes of a game, enabling human game designers to create enjoyable games more quickly.

A visualization of the Variational Transformer Network (VTN) model, which is able to extract meaningful relationships between the layout elements (paragraphs, tables, images, etc.) in order to generate realistic synthetic documents (e.g., with better alignment and margins).

Other ML algorithms have been used to evaluate the design space of computer architectural decisions for ML accelerator chips themselves. We’ve also shown that ML can be used to quickly create chip placements for ASIC designs that are better than layouts generated by human experts and can be generated in a matter of hours instead of weeks. This reduces the fixed engineering costs of chips and lowers the barrier to quickly creating specialized hardware for different applications. We’ve successfully used this automated placement approach in the design of our upcoming TPU-v5 chip.

Such exploratory ML approaches have also been applied to materials discovery. In a collaboration between Google Research and Caltech, several ML models, combined with a modified inkjet printer and a custom-built microscope, were able to rapidly search over hundreds of thousands of possible materials to hone in on 51 previously uncharacterized three-metal oxide materials with promising properties for applications in areas like battery technology and electrolysis of water.

These automated design space exploration approaches can help accelerate many scientific fields, especially when the entire experimental loop of generating the experiment and evaluating the result can all be done in an automated or mostly-automated manner. I expect to see this approach applied to good effect in many more areas in the coming years.

Application to Health

In addition to advancing basic science, ML can also drive advances in medicine and human health more broadly. The idea of leveraging advances in computer science in health is nothing new — in fact some of my own early experiences were in developing software to help analyze epidemiological data. But ML opens new doors, raises new opportunities, and yes, poses new challenges.

Take for example the field of genomics. Computing has been important to genomics since its inception, but ML adds new capabilities and disrupts old paradigms. When Google researchers began working in this area, the idea of using deep learning to help infer genetic variants from sequencer output was considered far-fetched by many experts. Today, this ML approach is considered state-of-the-art. But the future holds an even more important role for ML — genomics companies are developing new sequencing instruments that are more accurate and faster, but also present new inference challenges. Our release of open-source software DeepConsensus and, in collaboration with UCSC, PEPPER-DeepVariant, supports these new instruments with cutting-edge informatics. We hope that more rapid sequencing can lead to near term applicability with impact for real patients.

A schematic of the Transformer architecture for DeepConsensus, which corrects sequencing errors to improve yield and correctness.

There are other opportunities to use ML to accelerate our use of genomic information for personalized health outside of processing the sequencer data. Large biobanks of extensively phenotyped and sequenced individuals can revolutionize how we understand and manage genetic predisposition to disease. Our ML-based phenotyping method improves the scalability of converting large imaging and text datasets into phenotypes usable for genetic association studies, and our DeepNull method better leverages large phenotypic data for genetic discovery. We are happy to release both as open-source methods for the scientific community.

The process for generating large-scale quantification of anatomical and disease traits for combination with genomic data in Biobanks.

Just as ML helps us see hidden characteristics of genomics data, it can help us discover new information and glean new insights from other health data types as well. Diagnosis of disease is often about identifying a pattern, quantifying a correlation, or recognizing a new instance of a larger class — all tasks at which ML excels. Google researchers have used ML to tackle a wide range of such problems, but perhaps none of these has progressed farther than the applications of ML to medical imaging.

In fact, Google’s 2016 paper describing the application of deep learning to the screening for diabetic retinopathy, was selected by the editors of the Journal of the American Medical Association (JAMA) as one of the top 10 most influential papers of the decade — not just the most influential papers on ML and health, the most influential JAMA papers of the decade overall. But the strength of our research doesn’t end at contributions to the literature, but extends to our ability to build systems operating in the real world. Through our global network of deployment partners, this same program has helped screen tens of thousands of patients in India, Thailand, Germany and France who might otherwise have been untested for this vision-threatening disease.

We expect to see this same pattern of assistive ML systems deployed to improve breast cancer screening, detect lung cancer, accelerate radiotherapy treatments for cancer, flag abnormal X-rays, and stage prostate cancer biopsies. Each domain presents new opportunities to be helpful. ML-assisted colonoscopy procedures are a particularly interesting example of going beyond the basics. Colonoscopies are not just used to diagnose colon cancer — the removal of polyps during the procedure are the front line of halting disease progression and preventing serious illness. In this domain we’ve demonstrated that ML can help ensure doctors don’t miss polyps, can help detect elusive polyps, and can add new dimensions of quality assurance, like coverage mapping through the application of simultaneous localization and mapping techniques. In collaboration with Shaare Zedek Medical Center in Jerusalem, we’ve shown these systems can work in real time, detecting an average of one polyp per procedure that would have otherwise been missed, with fewer than four false alarms per procedure.

Sample chest X-rays (CXR) of true and false positives, and true and false negatives for (A) general abnormalities, (B) tuberculosis, and (C) COVID-19. On each CXR, red outlines indicate areas on which the model focused to identify abnormalities (i.e., the class activation map), and yellow outlines refer to regions of interest identified by a radiologist.

Another ambitious healthcare initiative, Care Studio, uses state-of-the-art ML and advanced NLP techniques to analyze structured data and medical notes, presenting clinicians with the most relevant information at the right time — ultimately helping them deliver more proactive and accurate care.

As important as ML may be to expanding access and improving accuracy in the clinical setting, we see a new equally important trend emerging: ML applied to help people in their daily health and well-being. Our everyday devices have powerful sensors that can help democratize health metrics and information so people can make more informed decisions about their health. We’ve already seen launches that enable a smartphone camera to assess heart rate and respiratory rate to help users without additional hardware, and Nest Hub devices that support contactless sleep sensing and allow users to better understand their nighttime wellness. We’ve seen that we can, on the one hand, significantly improve speech recognition quality for disordered speech in our own ASR systems, and on the other, use ML to help recreate the voice of those with speech impairments, empowering them to communicate in their own voice. ML enabled smartphones that help people better research emerging skin conditions or help those with limited vision go for a jog, seem to be just around the corner. These opportunities offer a future too bright to ignore.

The custom ML model for contactless sleep sensing efficiently processes a continuous stream of 3D radar tensors (summarizing activity over a range of distances, frequencies, and time) to automatically compute probabilities for the likelihood of user presence and wakefulness (awake or asleep).

ML Applications for the Climate Crisis

Another realm of paramount importance is climate change, which is an incredibly urgent threat for humanity. We need to all work together to bend the curve of harmful emissions to ensure a safe and prosperous future. Better information about the climate impact of different choices can help us tackle this challenge in a number of different ways.

To this end, we recently rolled out eco-friendly routing in Google Maps, which we estimate will save about 1 million tons of CO2 emissions per year (the equivalent of removing more than 200,000 cars from the road). A recent case study shows that using Google Maps directions in Salt Lake City results in both faster and more emissions-friendly routing, which saves 1.7% of CO2 emissions and 6.5% travel time. In addition, making our Maps products smarter about electric vehicles can help alleviate range anxiety, encouraging people to switch to emissions-free vehicles. We are also working with multiple municipalities around the world to use aggregated historical traffic data to help suggest improved traffic light timing settings, with an early pilot study in Israel and Brazil showing a 10-20% reduction in fuel consumption and delay time at the examined intersections.

With eco-friendly routing, Google Maps will show you the fastest route and the one that’s most fuel-efficient — so you can choose whichever one works best for you.

On a longer time scale, fusion holds promise as a game-changing renewable energy source. In a long-standing collaboration with TAE Technologies, we have used ML to help maintain stable plasmas in their fusion reactor by suggesting settings of the more than 1000 relevant control parameters. With our collaboration, TAE achieved their major goals for their Norman reactor, which brings us a step closer to the goal of breakeven fusion. The machine maintains a stable plasma at 30 million Kelvin (don’t touch!) for 30 milliseconds, which is the extent of available power to its systems. They have completed a design for an even more powerful machine, which they hope will demonstrate the conditions necessary for breakeven fusion before the end of the decade.

We’re also expanding our efforts to address wildfires and floods, which are becoming more common (like millions of Californians, I’m having to adapt to having a regular “fire season”). Last year, we launched a wildfire boundary map powered by satellite data to help people in the U.S. easily understand the approximate size and location of a fire — right from their device. Building on this, we’re now bringing all of Google’s wildfire information together and launching it globally with a new layer on Google Maps. We have been applying graph optimization algorithms to help optimize fire evacuation routes to help keep people safe in the presence of rapidly advancing fires. In 2021, our Flood Forecasting Initiative expanded its operational warning systems to cover 360 million people, and sent more than 115 million notifications directly to the mobile devices of people at risk from flooding, more than triple our outreach in the previous year. We also deployed our LSTM-based forecast models and the new Manifold inundation model in real-world systems for the first time, and shared a detailed description of all components of our systems.

The wildfire layer in Google Maps provides people with critical, up-to-date information in an emergency.

We’re also working hard on our own set of sustainability initiatives. Google was the first major company to become carbon neutral in 2007. We were also the first major company to match our energy use with 100 percent renewable energy in 2017. We operate the cleanest global cloud in the industry, and we’re the world’s largest corporate purchaser of renewable energy. Further, in 2020 we became the first major company to make a commitment to operate on 24/7 carbon-free energy in all our data centers and campuses worldwide. This is far more challenging than the traditional approach of matching energy usage with renewable energy, but we’re working to get this done by 2030. Carbon emission from ML model training is a concern for the ML community, and we have shown that making good choices about model architecture, datacenter, and ML accelerator type can reduce the carbon footprint of training by ~100-1000x.


Trend 5: Deeper and Broader Understanding of ML
As ML is used more broadly across technology products and society more generally, it is imperative that we continue to develop new techniques to ensure that it is applied fairly and equitably, and that it benefits all people and not just select subsets. This is a major focus for our Responsible AI and Human-Centered Technology research group and an area in which we conduct research on a variety of responsibility-related topics.

One area of focus is recommendation systems that are based on user activity in online products. Because these recommendation systems are often composed of multiple distinct components, understanding their fairness properties often requires insight into individual components as well as how the individual components behave when combined together. Recent work has helped to better understand these relationships, revealing ways to improve the fairness of both individual components and the overall recommendation system. In addition, when learning from implicit user activity, it is also important for recommendation systems to learn in an unbiased manner, since the straightforward approach of learning from items that were shown to previous users exhibits well-known forms of bias. Without correcting for such biases, for example, items that were shown in more prominent positions to users tend to get recommended to future users more often.

As in recommendation systems, surrounding context is important in machine translation. Because most machine translation systems translate individual sentences in isolation, without additional surrounding context, they can often reinforce biases related to gender, age or other areas. In an effort to address some of these issues, we have a long-standing line of research on reducing gender bias in our translation systems, and to help the entire translation community, last year we released a dataset to study gender bias in translation based on translations of Wikipedia biographies.

Another common problem in deploying machine learning models is distributional shift: if the statistical distribution of data on which the model was trained is not the same as that of the data the model is given as input, the model’s behavior can sometimes be unpredictable. In recent work, we employ the Deep Bootstrap framework to compare the real world, where there is finite training data, to an “ideal world”, where there is infinite data. Better understanding of how a model behaves in these two regimes (real vs. ideal) can help us develop models that generalize better to new settings and exhibit less bias towards fixed training datasets.

Although work on ML algorithms and model development gets significant attention, data collection and dataset curation often gets less. But this is an important area, because the data on which an ML model is trained can be a potential source of bias and fairness issues in downstream applications. Analyzing such data cascades in ML can help identify the many places in the lifecycle of an ML project that can have substantial influence on the outcomes. This research on data cascades has led to evidence-backed guidelines for data collection and evaluation in the revised PAIR Guidebook, aimed at ML developers and designers.

Arrows of different color indicate various types of data cascades, each of which typically originate upstream, compound over the ML development process, and manifest downstream.

The general goal of better understanding data is an important part of ML research. One thing that can help is finding and investigating anomalous data. We have developed methods to better understand the influence that particular training examples can have on an ML model, since mislabeled data or other similar issues can have outsized impact on the overall model behavior. We have also built the Know Your Data tool to help ML researchers and practitioners better understand properties of their datasets, and last year we created a case study of how to use the Know Your Data tool to explore issues like gender bias and age bias in a dataset.

A screenshot from Know Your Data showing the relationship between words that describe attractiveness and gendered words. For example, “attractive” and “male/man/boy” co-occur 12 times, but we expect ~60 times by chance (the ratio is 0.2x). On the other hand, “attractive” and “female/woman/girl” co-occur 2.62 times more than chance.

Understanding dynamics of benchmark dataset usage is also important, given the central role they play in the organization of ML as a field. Although studies of individual datasets have become increasingly common, the dynamics of dataset usage across the field have remained underexplored. In recent work, we published the first large scale empirical analysis of dynamics of dataset creation, adoption, and reuse. This work offers insights into pathways to enable more rigorous evaluations, as well as more equitable and socially informed research.

Creating public datasets that are more inclusive and less biased is an important way to help improve the field of ML for everyone. In 2016, we released the Open Images dataset, a collection of ~9 million images annotated with image labels spanning thousands of object categories and bounding box annotations for 600 classes. Last year, we introduced the More Inclusive Annotations for People (MIAP) dataset in the Open Images Extended collection. The collection contains more complete bounding box annotations for the person class hierarchy, and each annotation is labeled with fairness-related attributes, including perceived gender presentation and perceived age range. With the increasing focus on reducing unfair bias as part of responsible AI research, we hope these annotations will encourage researchers already leveraging the Open Images dataset to incorporate fairness analysis in their research.

Because we also know that our teams are not the only ones creating datasets that can improve machine learning, we have built Dataset Search to help users discover new and useful datasets, wherever they might be on the Web.

Tackling various forms of abusive behavior online, such as toxic language, hate speech, and misinformation, is a core priority for Google. Being able to detect such forms of abuse reliably, efficiently, and at scale is of critical importance both to ensure that our platforms are safe and also to avoid the risk of reproducing such negative traits through language technologies that learn from online discourse in an unsupervised fashion. Google has pioneered work in this space through the Perspective API tool, but the nuances involved in detecting toxicity at scale remains a complex problem. In recent work, in collaboration with various academic partners, we introduced a comprehensive taxonomy to reason about the changing landscape of online hate and harassment. We also investigated how to detect covert forms of toxicity, such as microaggressions, that are often ignored in online abuse interventions, studied how conventional approaches to deal with disagreements in data annotations of such subjective concepts might marginalize minority perspectives, and proposed a new disaggregated modeling approach that uses a multi-task framework to tackle this issue. Furthermore, through qualitative research and network-level content analysis, Google’s Jigsaw team, in collaboration with researchers at George Washington University, studied how hate clusters spread disinformation across social media platforms.

Another potential concern is that ML language understanding and generation models can sometimes also produce results that are not properly supported by evidence. To confront this problem in question answering, summarization, and dialog, we developed a new framework for measuring whether results can be attributed to specific sources. We released annotation guidelines and demonstrated that they can be reliably used in evaluating candidate models.

Interactive analysis and debugging of models remains key to responsible use of ML. We have updated our Language Interpretability Tool with new capabilities and techniques to advance this line of work, including support for image and tabular data, a variety of features carried over from our previous work on the What-If Tool, and built-in support for fairness analysis through the technique of Testing with Concept Activation Vectors. Interpretability and explainability of ML systems more generally is also a key part of our Responsible AI vision; in collaboration with DeepMind, we made headway in understanding the acquisition of human chess concepts in the self-trained AlphaZero chess system.

We are also working hard to broaden the perspective of Responsible AI beyond western contexts. Our recent research examines how various assumptions of conventional algorithmic fairness frameworks based on Western institutions and infrastructures may fail in non-Western contexts and offers a pathway for recontextualizing fairness research in India along several directions. We are actively conducting survey research across several continents to better understand perceptions of and preferences regarding AI. Western framing of algorithmic fairness research tends to focus on only a handful of attributes, thus biases concerning non-Western contexts are largely ignored and empirically under-studied. To address this gap, in collaboration with the University of Michigan, we developed a weakly supervised method to robustly detect lexical biases in broader geo-cultural contexts in NLP models that reflect human judgments of offensive and inoffensive language in those geographic contexts.

Furthermore, we have explored applications of ML to contexts valued in the Global South, including developing a proposal for farmer-centered ML research. Through this work, we hope to encourage the field to be thoughtful about how to bring ML-enabled solutions to smallholder farmers in ways that will improve their lives and their communities.

Involving community stakeholders at all stages of the ML pipeline is key to our efforts to develop and deploy ML responsibly and keep us focused on tackling the problems that matter most. In this vein, we held a Health Equity Research Summit among external faculty, non-profit organization leads, government and NGO representatives, and other subject matter experts to discuss how to bring more equity into the entire ML ecosystem, from the way we approach problem-solving to how we assess the impact of our efforts.

Community-based research methods have also informed our approach to designing for digital wellbeing and addressing racial equity issues in ML systems, including improving our understanding of the experience of Black Americans using ASR systems. We are also listening to the public more broadly to learn how sociotechnical ML systems could help during major life events, such as by supporting family caregiving.

As ML models become more capable and have impact in many domains, the protection of the private information used in ML continues to be an important focus for research. Along these lines, some of our recent work addresses privacy in large models, both highlighting that training data can sometimes be extracted from large models and pointing to how privacy can be achieved in large models, e.g., as in differentially private BERT. In addition to the work on federated learning and analytics, mentioned above, we have also been enhancing our toolbox with other principled and practical ML techniques for ensuring differential privacy, for example private clustering, private personalization, private matrix completion, private weighted sampling, private quantiles, private robust learning of halfspaces, and in general, sample-efficient private PAC learning. Moreover, we have been expanding the set of privacy notions that can be tailored to different applications and threat models, including label privacy and user versus item level privacy.

A visual illustration of the differentially private clustering algorithm.


Recognizing the value of open datasets to the general advancement of ML and related fields of research, we continue to grow our collection of open source datasets and resources and expand our global index of open datasets in Google Dataset Search. This year, we have released a number of datasets and tools across a range of research areas:

Datasets & Tools Description
AIST++ 3D keypoints with corresponding images for dance motions covering 10 dance genres
AutoFlow 40k image pairs with ground truth optical flow
C4_200M A 200 million sentence synthetic dataset for grammatical error correction
CIFAR-5M Dataset of ~6 million synthetic CIFAR-10–like images (RGB 32 x 32 pix)
Crisscrossed Captions Set of semantic similarity ratings for the MS-COCO dataset
Disfl-QA Dataset of contextual disfluencies for information seeking
Distilled Datasets Distilled datasets from CIFAR-10, CIFAR-100, MNIST, Fashion-MNIST, and SVHN
EvolvingRL 1000 top performing RL algorithms discovered through algorithm evolution
GoEmotions A human-annotated dataset of 58k Reddit comments labeled with 27 emotion categories
H01 Dataset 1.4 petabyte browsable reconstruction of the human cortex
Know Your Data Tool for understanding biases in a dataset
Lens Flare 5000 high-quality RGB images of typical lens flare
More Inclusive Annotations for People (MIAP) Improved bounding box annotations for a subset of the person class in the Open Images dataset
Mostly Basic Python Problems 1000 Python programming problems, incl. task description, code solution & test cases
NIH ChestX-ray14 dataset labels Expert labels for a subset of the NIH ChestX-ray14 dataset
Open Buildings Locations and footprints of 516 million buildings with coverage across most of Africa
Optical Polarization from Curie 5GB of optical polarization data from the Curie submarine cable
Readability Scroll Scroll interactions of ~600 participants reading texts from the OneStopEnglish corpus
RLDS Tools to store, retrieve & manipulate episodic data for reinforcement learning
Room-Across-Room (RxR) Multilingual dataset for vision-and-language navigation in English, Hindi and Telugu
Soft Attributes ~6k sets of movie titles annotated with single English soft attributes
TimeDial Dataset of multiple choice span-filling tasks for temporal commonsense reasoning in dialog
ToTTo English table-to-text generation dataset with a controlled text generation task
Translated Wikipedia Biographies Dataset for analysis of common gender errors in NMT for English, Spanish and German
UI Understanding Data for UIBert Datasets for two UI understanding tasks, AppSim & RefExp
WikiFact Wikipedia & WikiData–based dataset to train relationship classifiers and fact extraction models
WIT Wikipedia-based Image Text dataset for multimodal multilingual ML

Research Community Interaction
To realize our goal for a more robust and comprehensive understanding of ML and related technologies, we actively engage with the broader research community. In 2021, we published over 750 papers, nearly 600 of which were presented at leading research conferences. Google Research sponsored over 150 conferences, and Google researchers contributed directly by serving on program committees and organizing workshops, tutorials and numerous other activities aimed at collectively advancing the field. To learn more about our contributions to some of the larger research conferences this year, please see our recent conference blog posts. In addition, we hosted 19 virtual workshops (like the 2021 Quantum Summer Symposium), which allowed us to further engage with the academic community by generating new ideas and directions for the research field and advancing research initiatives.

In 2021, Google Research also directly supported external research with $59M in funding, including $23M through Research programs to faculty and students, and $20M in university partnerships and outreach. This past year, we introduced new funding and collaboration programs that support academics all over the world who are doing high impact research. We funded 86 early career faculty through our Research Scholar Program to support general advancements in science, and funded 34 faculty through our Award for Inclusion Research Program who are doing research in areas like accessibility, algorithmic fairness, higher education and collaboration, and participatory ML. In addition to the research we are funding, we welcomed 85 faculty and post-docs, globally, through our Visiting Researcher program, to come to Google and partner with us on exciting ideas and shared research challenges. We also selected a group of 74 incredibly talented PhD student researchers to receive Google PhD Fellowships and mentorship as they conduct their research.

As part of our ongoing racial equity commitments, making computer science (CS) research more inclusive continues to be a top priority for us. In 2021, we continued expanding our efforts to increase the diversity of Ph.D. graduates in computing. For example, the CS Research Mentorship Program (CSRMP), an initiative by Google Research to support students from historically marginalized groups (HMGs) in computing research pathways, graduated 590 mentees, 83% of whom self-identified as part of an HMG, who were supported by 194 Google mentors — our largest group to date! In October, we welcomed 35 institutions globally leading the way to engage 3,400+ students in computing research as part of the 2021 exploreCSR cohort. Since 2018, this program has provided faculty with funding, community, evaluation and connections to Google researchers in order to introduce students from HMGs to the world of CS research. We are excited to expand this program to more international locations in 2022.

We also continued our efforts to fund and partner with organizations to develop and support new pathways and approaches to broadening participation in computing research at scale. From working with alliances like the Computing Alliance of Hispanic-Serving Institutions (CAHSI) and CMD-IT Diversifying LEAdership in the Professoriate (LEAP) Alliance to partnering with university initiatives like UMBC’s Meyerhoff Scholars, Cornell University’s CSMore, Northeastern University’s Center for Inclusive Computing, and MIT’s MEnTorEd Opportunities in Research (METEOR), we are taking a community-based approach to materially increase the representation of marginalized groups in computing research.

Other Work
In writing these retrospectives, I try to focus on new research work that has happened (mostly) in the past year while also looking ahead. In past years’ retrospectives, I’ve tried to be more comprehensive, but this time I thought it could be more interesting to focus on just a few themes. We’ve also done great  work in many other research areas that don’t fit neatly into these themes. If you’re interested, I encourage you to check out our research publications by area below or by year (and if you’re interested in quantum computing, our Quantum team recently wrote a retrospective of their work in 2021):

Algorithms and Theory Hardware and Architecture Networking
Data Management Human-Computer Interaction and Visualization Quantum Computing
Data Mining Information Retrieval and the Web Responsible AI
Distributed Systems & Parallel Computing Machine Intelligence Robotics
Economics & Electronic Commerce Machine Perception Security, Privacy and Abuse Prevention
Education Innovation Machine Translation Software Engineering
General Science Mobile Systems Software Systems
Health and Bioscience Natural Language Processing Speech Processing

Research is often a multi-year journey to real-world impact. Early stage research work that happened a few years ago is now having a dramatic impact on Google’s products and across the world. Investments in ML hardware accelerators like TPUs and in software frameworks like TensorFlow and JAX have borne fruit. ML models are increasingly prevalent in many different products and features at Google because their power and ease of expression streamline experimentation and productionization of ML models in performance-critical environments. Research into model architectures to create Seq2Seq, Inception, EfficientNet, and Transformer or algorithmic research like batch normalization and distillation is driving progress in the fields of language understanding, vision, speech, and others. Basic capabilities like better language and visual understanding and speech recognition can be transformational, and as a result, these sorts of models are widely deployed for a wide variety of problems in many of our products including Search, Assistant, Ads, Cloud, Gmail, Maps, YouTube, Workspace, Android, Pixel, Nest, and Translate.

These are truly exciting times in machine learning and computer science. Continued improvement in computers’ ability to understand and interact with the world around them through language, vision, and sound opens up entire new frontiers of how computers can help people accomplish things in the world. The many examples of progress along the five themes outlined in this post are waypoints in a long-term journey!

Thanks to Alison Carroll, Alison Lentz, Andrew Carroll, Andrew Tomkins, Avinatan Hassidim, Azalia Mirhoseini, Barak Turovsky, Been Kim, Blaise Aguera y Arcas, Brennan Saeta, Brian Rakowski, Charina Chou, Christian Howard, Claire Cui, Corinna Cortes, Courtney Heldreth, David Patterson, Dipanjan Das, Ed Chi, Eli Collins, Emily Denton, Fernando Pereira, Genevieve Park, Greg Corrado, Ian Tenney, Iz Conroy, James Wexler, Jason Freidenfelds, John Platt, Katherine Chou, Kathy Meier-Hellstern, Kyle Vandenberg, Lauren Wilcox, Lizzie Dorfman, Marian Croak, Martin Abadi, Matthew Flegal, Meredith Morris, Natasha Noy, Negar Saei, Neha Arora, Paul Muret, Paul Natsev, Quoc Le, Ravi Kumar, Rina Panigrahy, Sanjiv Kumar, Sella Nevo, Slav Petrov, Sreenivas Gollapudi, Tom Duerig, Tom Small, Vidhya Navalpakkam, Vincent Vanhoucke, Vinodkumar Prabhakaran, Viren Jain, Yonghui Wu, Yossi Matias, and Zoubin Ghahramani for helpful feedback and contributions to this post, and to the entire Research and Health communities at Google for everyone’s contributions towards this work.


Alice, Bob, and the average shadow of a cube


A Scalable Approach for Partially Local Federated Learning

Federated learning enables users to train a model without sending raw data to a central server, thus avoiding the collection of privacy-sensitive data. Often this is done by learning a single global model for all users, even though the users may differ in their data distributions. For example, users of a mobile keyboard application may collaborate to train a suggestion model but have different preferences for the suggestions. This heterogeneity has motivated algorithms that can personalize a global model for each user.

However, in some settings privacy considerations may prohibit learning a fully global model. Consider models with user-specific embeddings, such as matrix factorization models for recommender systems. Training a fully global federated model would involve sending user embedding updates to a central server, which could potentially reveal the preferences encoded in the embeddings. Even for models without user-specific embeddings, having some parameters be completely local to user devices would reduce server-client communication and responsibly personalize those parameters to each user.

Left: A matrix factorization model with a user matrix P and items matrix Q. The user embedding for a user u (Pu) and item embedding for item i (Qi) are trained to predict the user’s rating for that item (Rui). Right: Applying federated learning approaches to learn a global model can involve sending updates for Pu to a central server, potentially leaking individual user preferences.

In “Federated Reconstruction: Partially Local Federated Learning”, presented at NeurIPS 2021, we introduce an approach that enables scalable partially local federated learning, where some model parameters are never aggregated on the server. For matrix factorization, this approach trains a recommender model while keeping user embeddings local to each user device. For other models, this approach trains a portion of the model to be completely personal for each user while avoiding communication of these parameters. We successfully deployed partially local federated learning to Gboard, resulting in better recommendations for hundreds of millions of keyboard users. We’re also releasing a TensorFlow Federated tutorial demonstrating how to use Federated Reconstruction.

Federated Reconstruction
Previous approaches for partially local federated learning used stateful algorithms, which require user devices to store a state across rounds of federated training. Specifically, these approaches required devices to store local parameters across rounds. However, these algorithms tend to degrade in large-scale federated learning settings. In these cases, the majority of users do not participate in training, and users who do participate likely only do so once, resulting in a state that is rarely available and can get stale across rounds. Also, all users who do not participate are left without trained local parameters, preventing practical applications.

Federated Reconstruction is stateless and avoids the need for user devices to store local parameters by reconstructing them whenever needed. When a user participates in training, before updating any globally aggregated model parameters, they randomly initialize and train their local parameters using gradient descent on local data with global parameters frozen. They can then calculate updates to global parameters with local parameters frozen. A round of Federated Reconstruction training is depicted below.

Models are partitioned into global and local parameters. For each round of Federated Reconstruction training: (1) The server sends the current global parameters g to each user i; (2) Each user i freezes g and reconstructs their local parameters li; (3) Each user i freezes li and updates g to produce gi; (4) Users’ gi are averaged to produce the global parameters for the next round. Steps (2) and (3) generally use distinct parts of the local data.

This simple approach avoids the challenges of previous methods. It does not assume users have a state from previous rounds of training, enabling large-scale training, and local parameters are always freshly reconstructed, preventing staleness. Users unseen during training can still get trained models and perform inference by simply reconstructing local parameters using local data.

Federated Reconstruction trains better performing models for unseen users compared to other approaches. For a matrix factorization task with unseen users, the approach significantly outperforms both centralized training and baseline Federated Averaging.

RMSE ↓ Accuracy ↑
Centralized 1.36 40.8%
FedAvg .934 40.0%
FedRecon (this work) .907 43.3%
Root-mean-square-error (lower is better) and accuracy for a matrix factorization task with unseen users. Centralized training and Federated Averaging (FedAvg) both reveal privacy-sensitive user embeddings to a central server, while Federated Reconstruction (FedRecon) avoids this.

These results can be explained via a connection to meta learning (i.e., learning to learn); Federated Reconstruction trains global parameters that lead to fast and accurate reconstruction of local parameters for unseen users. That is, Federated Reconstruction is learning to learn local parameters. In practice, we observe that just one gradient descent step can yield successful reconstruction, even for models with about one million local parameters.

Federated Reconstruction also provides a way to personalize models for heterogeneous users while reducing communication of model parameters — even for models without user-specific embeddings. To evaluate this, we apply Federated Reconstruction to personalize a next word prediction language model and observe a substantial increase in performance, attaining accuracy on par with other personalization methods despite reduced communication. Federated Reconstruction also outperforms other personalization methods when executed at a fixed communication level.

Accuracy ↑ Communication ↓
FedYogi 24.3% Whole Model
FedYogi + Finetuning 30.8% Whole Model
FedRecon (this work) 30.7% Partial Model
Accuracy and server-client communication for a next word prediction task without user-specific embeddings. FedYogi communicates all model parameters, while FedRecon avoids this.

Real-World Deployment in Gboard
To validate the practicality of Federated Reconstruction in large-scale settings, we deployed the algorithm to Gboard, a mobile keyboard application with hundreds of millions of users. Gboard users use expressions (e.g., GIFs, stickers) to communicate with others. Users have highly heterogeneous preferences for these expressions, making the setting a good fit for using matrix factorization to predict new expressions a user might want to share.

Gboard users can communicate with expressions, preferences for which are highly personal.

We trained a matrix factorization model over user-expression co-occurrences using Federated Reconstruction, keeping user embeddings local to each Gboard user. We then deployed the model to Gboard users, leading to a 29.3% increase in click-through-rate for expression recommendations. Since most Gboard users were unseen during federated training, Federated Reconstruction played a key role in this deployment.

Further Explorations
We’ve presented Federated Reconstruction, a method for partially local federated learning. Federated Reconstruction enables personalization to heterogeneous users while reducing communication of privacy-sensitive parameters. We scaled the approach to Gboard in alignment with our AI Principles, improving recommendations for hundreds of millions of users.

For a technical walkthrough of Federated Reconstruction for matrix factorization, check out the TensorFlow Federated tutorial. We’ve also released general-purpose TensorFlow Federated libraries and open-source code for running experiments.

Karan Singhal, Hakim Sidahmed, Zachary Garrett, Shanshan Wu, Keith Rush, and Sushant Prakash co-authored the paper. Thanks to Wei Li, Matt Newton, and Yang Lu for their partnership on Gboard deployment. We’d also like to thank Brendan McMahan, Lin Ning, Zachary Charles, Warren Morningstar, Daniel Ramage, Jakub Konecný, Alex Ingerman, Blaise Agüera y Arcas, Jay Yagnik, Bradley Green, and Ewa Dominowska for their helpful comments and support.


Training Machine Learning Models More Efficiently with Dataset Distillation

For a machine learning (ML) algorithm to be effective, useful features must be extracted from (often) large amounts of training data. However, this process can be made challenging due to the costs associated with training on such large datasets, both in terms of compute requirements and wall clock time. The idea of distillation plays an important role in these situations by reducing the resources required for the model to be effective. The most widely known form of distillation is model distillation (a.k.a. knowledge distillation), where the predictions of large, complex teacher models are distilled into smaller models.

An alternative option to this model-space approach is dataset distillation [1, 2], in which a large dataset is distilled into a synthetic, smaller dataset. Training a model with such a distilled dataset can reduce the required memory and compute. For example, instead of using all 50,000 images and labels of the CIFAR-10 dataset, one could use a distilled dataset consisting of only 10 synthesized data points (1 image per class) to train an ML model that can still achieve good performance on the unseen test set.

Top: Natural (i.e., unmodified) CIFAR-10 images. Bottom: Distilled dataset (1 image per class) on CIFAR-10 classification task. Using only these 10 synthetic images as training data, a model can achieve test set accuracy of ~51%.

In “Dataset Meta-Learning from Kernel Ridge Regression”, published in ICLR 2021, and “Dataset Distillation with Infinitely Wide Convolutional Networks”, presented at NeurIPS 2021, we introduce two novel dataset distillation algorithms, Kernel Inducing Points (KIP) and Label Solve (LS), which optimize datasets using the loss function arising from kernel regression (a classical machine learning algorithm that fits a linear model to features defined through a kernel). Applying the KIP and LS algorithms, we obtain very efficient distilled datasets for image classification, reducing the datasets to 1, 10, or 50 data points per class while still obtaining state-of-the-art results on a number of benchmark image classification datasets. Additionally, we are also excited to release our distilled datasets to benefit the wider research community.

One of the key theoretical insights of deep neural networks (DNN) in recent years has been that increasing the width of DNNs results in more regular behavior that makes them easier to understand. As the width is taken to infinity, DNNs trained by gradient descent converge to the familiar and simpler class of models arising from kernel regression with respect to the neural tangent kernel (NTK), a kernel that measures input similarity by computing dot products of gradients of the neural network. Thanks to the Neural Tangents library, neural kernels for various DNN architectures can be computed in a scalable manner.

We utilized the above infinite-width limit theory of neural networks to tackle dataset distillation. Dataset distillation can be formulated as a two-stage optimization process: an “inner loop” that trains a model on learned data, and an “outer loop” that optimizes the learned data for performance on natural (i.e., unmodified) data. The infinite-width limit replaces the inner loop of training a finite-width neural network with a simple kernel regression. With the addition of a regularizing term, the kernel regression becomes a kernel ridge-regression (KRR) problem. This is a highly valuable outcome because the kernel ridge regressor (i.e., the predictor from the algorithm) has an explicit formula in terms of its training data (unlike a neural network predictor), which means that one can easily optimize the KRR loss function during the outer loop.

The original data labels can be represented by one-hot vectors, i.e., the true label is given a value of 1 and all other labels are given values of 0. Thus, an image of a cat would have the label “cat” assigned a 1 value, while the labels for “dog” and “horse” would be 0. The labels we use involve a subsequent mean-centering step, where we subtract the reciprocal of the number of classes from each component (so 0.1 for 10-way classification) so that the expected value of each label component across the dataset is normalized to zero.

While the labels for natural images appear in this standard form, the labels for our learned distilled datasets are free to be optimized for performance. Having obtained the kernel ridge regressor from the inner loop, the KRR loss function in the outer loop computes the mean-square error between the original labels of natural images and the labels predicted by the kernel ridge regressor. KIP optimizes the support data (images and possibly labels) by minimizing the KRR loss function through gradient-based methods. The Label Solve algorithm directly solves for the set of support labels that minimizes the KRR loss function, generating a unique dense label vector for each (natural) support image.

Example of labels obtained by label solving. Left and Middle: Sample images with possible labels listed below. The raw, one-hot label is shown in blue and the final LS generated dense label is shown in orange. Right: The covariance matrix between original labels and learned labels. Here, 500 labels were distilled from the CIFAR-10 dataset. A test accuracy of 69.7% is achieved using these labels for kernel ridge-regression.

Distributed Computation
For simplicity, we focus on architectures that consist of convolutional neural networks with pooling layers. Specifically, we focus on the so-called “ConvNet” architecture and its variants because it has been featured in other dataset distillation studies. We used a slightly modified version of ConvNet that has a simple architecture given by three blocks of convolution, ReLu, and 2×2 average pooling and then a final linear readout layer, with an additional 3×3 convolution and ReLu layer prepended (see our GitHub for precise details).

ConvNet architecture used in DC/DSA. Ours has an additional 3×3 Conv and ReLu prepended.

To compute the neural kernels needed in our work, we used the Neural Tangents library.

The first stage of this work, in which we applied KRR, focused on fully-connected networks, whose kernel elements are cheap to compute. But a hurdle facing neural kernels for models with convolutional layers plus pooling is that the computation of each kernel element between two images scales as the square of the number of input pixels (due to the capturing of pixel-pixel correlations by the kernel). So, for the second stage of this work, we needed to distribute the computation of the kernel elements and their gradients across many devices.

Distributed computation for large scale metalearning.

We invoke a client-server model of distributed computation in which a server distributes independent workloads to a large pool of client workers. A key part of this is to divide the backpropagation step in a way that is computationally efficient (explained in detail in the paper).

We accomplish this using the open-source tools Courier (part of DeepMind’s Launchpad), which allows us to distribute computations across GPUs working in parallel, and JAX, for which novel usage of the jax.vjp function enables computationally efficient gradients. This distributed framework allows us to utilize hundreds of GPUs per distillation of the dataset, for both the KIP and LS algorithms. Given the compute required for such experiments, we are releasing our distilled datasets to benefit the wider research community.

Our first set of distilled images above used KIP to distill CIFAR-10 down to 1 image per class while keeping the labels fixed. Next, in the below figure, we compare the test accuracy of training on natural MNIST images, KIP distilled images with labels fixed, and KIP distilled images with labels optimized. We highlight that learning the labels provides an effective, albeit mysterious benefit to distilling datasets. Indeed the resulting set of images provides the best test performance (for infinite-width networks) despite being less interpretable.

MNIST dataset distillation with trainable and non-trainable labels. Top: Natural MNIST data. Middle: Kernel Inducing Point distilled data with fixed labels. Bottom: Kernel Inducing Point distilled data with learned labels.

Our distilled datasets achieve state-of-the-art performance on benchmark image classification datasets, improving performance beyond previous state-of-the-art models that used convolutional architectures, Dataset Condensation (DC) and Dataset Condensation with Differentiable Siamese Augmentation (DSA). In particular, for CIFAR-10 classification tasks, a model trained on a dataset consisting of only 10 distilled data entries (1 image / class, 0.02% of the whole dataset) achieves a 64% test set accuracy. Here, learning labels and an additional image preprocessing step leads to a significant increase in performance beyond the 50% test accuracy shown in our first figure (see our paper for details). With 500 images (50 images / class, 1% of the whole dataset), the model reaches 80% test set accuracy. While these numbers are with respect to neural kernels (using the KRR infinite width limit), these distilled datasets can be used to train finite-width neural networks as well. In particular, for 10 data points on CIFAR-10, a finite-width ConvNet neural network achieves 50% test accuracy with 10 images and 68% test accuracy using 500 images, which are still state-of-the-art results. We provide a simple Colab notebook demonstrating this transfer to a finite-width neural network.

Dataset distillation using Kernel Inducing Points (KIP) with a convolutional architecture outperforms prior state-of-the-art models (DC/DSA) on all benchmark settings on image classification tasks. Label Solve (LS, middle columns) while only distilling information in the labels could often (e.g. CIFAR-10 10, 50 data points per class) outperform prior state-of-the-art models as well.

In some cases, our learned datasets are more effective than a natural dataset one hundred times larger in size.

We believe that our work on dataset distillation opens up many interesting future directions. For instance, our algorithms KIP and LS have demonstrated the effectiveness of using learned labels, an area that remains relatively underexplored. Furthermore, we expect that utilizing efficient kernel approximation methods can help to reduce computational burden and scale up to larger datasets. We hope this work encourages researchers to explore other applications of dataset distillation, including neural architecture search and continual learning, and even potential applications to privacy.

Anyone interested in the KIP and LS learned datasets for further analysis is encouraged to check out our papers [ICLR 2021, NeurIPS 2021] and open-sourced code and datasets available on Github.

This project was done in collaboration with Zhourong Chen, Roman Novak and Lechao Xiao. We would like to acknowledge special thanks to Samuel S. Schoenholz, who proposed and helped develop the overall strategy for our distributed KIP learning methodology.

1Now at DeepMind.  


Interpretable Deep Learning for Time Series Forecasting

Multi-horizon forecasting, i.e. predicting variables-of-interest at multiple future time steps, is a crucial challenge in time series machine learning. Most real-world datasets have a time component, and forecasting the future can unlock great value. For example, retailers can use future sales to optimize their supply chain and promotions, investment managers are interested in forecasting the future prices of financial assets to maximize their performance, and healthcare institutions can use the number of future patient admissions to have sufficient personnel and equipment.

Deep neural networks (DNNs) have increasingly been used in multi-horizon forecasting, demonstrating strong performance improvements over traditional time series models. While many models (e.g., DeepAR, MQRNN) have focused on variants of recurrent neural networks (RNNs), recent improvements, including Transformer-based models, have used attention-based layers to enhance the selection of relevant time steps in the past beyond the inductive bias of RNNs – sequential ordered processing of information including. However, these often do not consider the different inputs commonly present in multi-horizon forecasting and either assume that all exogenous inputs are known into the future or neglect important static covariates.

Multi-horizon forecasting with static covariates and various time-dependent inputs.

Additionally, conventional time series models are controlled by complex nonlinear interactions between many parameters, making it difficult to explain how such models arrive at their predictions. Unfortunately, common methods to explain the behavior of DNNs have limitations. For example, post-hoc methods (e.g., LIME and SHAP) do not consider the order of input features. Some attention-based models are proposed with inherent interpretability for sequential data, primarily language or speech, but multi-horizon forecasting has many different types of inputs, not just language or speech. Attention-based models can provide insights into relevant time steps, but they cannot distinguish the importance of different features at a given time step. New methods are needed to tackle the heterogeneity of data in multi-horizon forecasting for high performance and to render these forecasts interpretable.

To that end, we announce “Temporal Fusion Transformers for Interpretable Multi-horizon Time Series Forecasting”, published in the International Journal of Forecasting, where we propose the Temporal Fusion Transformer (TFT), an attention-based DNN model for multi-horizon forecasting. TFT is designed to explicitly align the model with the general multi-horizon forecasting task for both superior accuracy and interpretability, which we demonstrate across various use cases.

Temporal Fusion Transformer
We design TFT to efficiently build feature representations for each input type (i.e., static, known, or observed inputs) for high forecasting performance. The major constituents of TFT (shown below) are:

  1. Gating mechanismsto skip over any unused components of the model (learned from the data), providing adaptive depth and network complexity to accommodate a wide range of datasets.
  2. Variable selection networksto select relevant input variables at each time step. While conventional DNNs may overfit to irrelevant features, attention-based variable selection can improve generalization by encouraging the model to anchor most of its learning capacity on the most salient features.
  3. Static covariate encodersintegrate static features to control how temporal dynamics are modeled. Static features can have an important impact on forecasts, e.g., a store location could have different temporal dynamics for sales (e.g., a rural store may see higher weekend traffic, but a downtown store may see daily peaks after working hours).
  4. Temporal processingto learn both long- and short-term temporal relationships from both observed and known time-varying inputs. A sequence-to-sequence layer is employed for local processing as the inductive bias it has for ordered information processing is beneficial, whereas long-term dependencies are captured using a novel interpretable multi-head attention block. This can cut the effective path length of information, i.e., any past time step with relevant information (e.g. sales from last year) can be focused on directly.
  5. Prediction intervals show quantile forecasts to determine the range of target values at each prediction horizon, which help users understand the distribution of the output, not just the point forecasts.
TFT inputs static metadata, time-varying past inputs and time-varying a priori known future inputs. Variable Selection is used for judicious selection of the most salient features based on the input. Gated information is added as a residual input, followed by normalization. Gated residual network (GRN) blocks enable efficient information flow with skip connections and gating layers. Time-dependent processing is based on LSTMs for local processing, and multi-head attention for integrating information from any time step.

Forecasting Performance
We compare TFT to a wide range of models for multi-horizon forecasting, including various deep learning models with iterative methods (e.g., DeepAR, DeepSSM, ConvTrans) and direct methods (e.g., LSTM Seq2Seq, MQRNN), as well as traditional models such as ARIMA, ETS, and TRMF. Below is a comparison to a truncated list of models.

Model Electricity Traffic Volatility Retail
ARIMA 0.154 (+180%) 0.223 (+135%)
ETS 0.102 (+85%) 0.236 (+148%)
DeepAR 0.075 (+36%) 0.161 (+69%) 0.050 (+28%) 0.574 (+62%)
Seq2Seq 0.067 (+22%) 0.105 (+11%) 0.042 (+7%) 0.411 (+16%)
MQRNN 0.077 (+40%) 0.117 (+23%) 0.042 (+7%) 0.379 (+7%)
TFT 0.055 0.095 0.039 0.354
P50 quantile losses (lower is better) for TFT vs. alternative models.

As shown above, TFT outperforms all benchmarks over a variety of datasets. This applies to both point forecasts and uncertainty estimates, with TFT yielding an average 7% lower P50 and 9% lower P90 losses, respectively, compared to the next best model.

Interpretability Use Cases
We demonstrate how TFT’s design allows for analysis of its individual components for enhanced interpretability with three use cases.

  • Variable Importance
    One can observe how different variables impact retail sales by observing their model weights. For example, the largest weights for static variables were the specific store and item, while the largest weights for future variables were promotion period and national holiday (shown below).

    Variable importance for the retail dataset. The 10th, 50th, and 90th percentiles of the variable selection weights are shown, with values larger than 0.1 in bold purple.
  • Persistent Temporal Patterns
    Visualizing persistent temporal patterns can help in understanding the time-dependent relationships present in a given dataset. We identify similar persistent patterns by measuring the contributions of features at fixed lags in the past forecasts at various horizons. Shown below, attention weights reveal the most important past time steps on which TFT bases its decisions.

    Persistent temporal patterns for the traffic dataset (𝛕 denotes the forecasting horizon) for the 10%, 50% and 90% quantile levels. Clear periodicity is observed with peaks being separated by ~24 hours, i.e., the model attends the most to the time steps that are at the same time of the day from past days, which is aligned with the expected daily traffic patterns.

    The above shows the attention weight patterns across time, indicating how TFT learns persistent temporal patterns without any hard-coding. Such capability can help build trust with users because the output confirms expected known patterns. Model developers can also use these towards model improvements, e.g., via specific feature engineering or data collection.

  • Identifying Significant Events
    Identifying sudden changes can be useful, as temporary shifts can occur due to the presence of significant events. TFT uses the distance between attention patterns at each point with the average pattern to identify the significant deviations. The figures below show that TFT can alter its attention between events — placing equal attention across past inputs when volatility is low, while attending more to sharp trend changes during high volatility periods.

    Event identification for S&P 500 realized volatility from 2002 through 2014.

    Significant deviations in attention patterns can be observed above around periods of high volatility, corresponding to the peaks observed in dist(t), distance between attention patterns (red line). We use a threshold to denote significant events, as highlighted in purple.

    Focusing on periods around the 2008 financial crisis, the bottom plot below zooms on midway through the significant event (evident from the increased attention on sharp trend changes), compared to the normal event in the top plot (where attention is equal over low volatility periods).

    Event identification for S&P 500 realized volatility, a zoom of the above on a period from 2004 and 2005.

    Event identification for S&P 500 realized volatility, a zoom of the above on a period from 2008 and 2009.

Real-World Impact
Finally, TFT has been used to help retail and logistics companies with demand forecasting by both improving forecasting accuracy and providing interpretability capabilities.

Additionally, TFT has potential applications for climate-related challenges: for example, reducing greenhouse gas emissions by balancing electricity supply and demand in real time, and improving the accuracy and interpretability of rainfall forecasting results.

We present a novel attention-based model for high-performance multi-horizon forecasting. In addition to improved performance across a range of datasets, TFT also contains specialized components for inherent interpretability — i.e., variable selection networks and interpretable multi-head attention. With three interpretability use-cases, we also demonstrate how these components can be used to extract insights on feature importance and temporal dynamics.

We gratefully acknowledge contributions of Bryan Lim, Nicolas Loeff, Minho Jin, Yaguang Li, and Andrew Moore.


A Fast WordPiece Tokenization System

Tokenization is a fundamental pre-processing step for most natural language processing (NLP) applications. It involves splitting text into smaller units called tokens (e.g., words or word segments) in order to turn an unstructured input string into a sequence of discrete elements that is suitable for a machine learning (ML) model. ln deep learning–based models (e.g., BERT), each token is mapped to an embedding vector to be fed into the model.

Tokenization in a typical deep learning model, like BERT.

A fundamental tokenization approach is to break text into words. However, using this approach, words that are not included in the vocabulary are treated as “unknown”. Modern NLP models address this issue by tokenizing text into subword units, which often retain linguistic meaning (e.g., morphemes). So, even though a word may be unknown to the model, individual subword tokens may retain enough information for the model to infer the meaning to some extent. One such subword tokenization technique that is commonly used and can be applied to many other NLP models is called WordPiece. Given text, WordPiece first pre-tokenizes the text into words (by splitting on punctuation and whitespaces) and then tokenizes each word into subword units, called wordpieces.

The WordPiece tokenization process with an example sentence.

In “Fast WordPiece Tokenization”, presented at EMNLP 2021, we developed an improved end-to-end WordPiece tokenization system that speeds up the tokenization process, reducing the overall model latency and saving computing resources. In comparison to traditional algorithms that have been used for decades, this approach reduces the complexity of the computation by an order of magnitude, resulting in significantly improved performance, up to 8x faster than standard approaches. The system has been applied successfully in a number of systems at Google and has been publicly released in TensorFlow Text.

Single-Word WordPiece Tokenization
WordPiece uses a greedy longest-match-first strategy to tokenize a single word — i.e., it iteratively picks the longest prefix of the remaining text that matches a word in the model’s vocabulary. This approach is known as maximum matching or MaxMatch, and has also been used for Chinese word segmentation since the 1980s. Yet despite its wide use in NLP for decades, it is still relatively computation intensive, with the commonly adopted MaxMatch approaches’ computation being quadratic with respect to the input word length (n). This is because two pointers are needed to scan over the input: one to mark a start position, and the other to search for the longest substring matching a vocabulary token at that position.

We propose an alternative to the MaxMatch algorithm for WordPiece tokenization, called LinMaxMatch, which has a tokenization time that is strictly linear with respect to n. First, we organize the vocabulary tokens in a trie (also called a prefix tree), where each trie edge is labeled by a character, and a tree path from the root to some node represents a prefix of some token in the vocabulary. In the figure below, nodes are depicted as circles and tree edges are black solid arrows. Given a trie, a vocabulary token can be located to match an input text by traversing from the root and following the trie edges to match the input character by character; this process is referred to as trie matching.

The figure below shows the trie created from the vocabulary consisting of “a”, “abcd”, “##b”, “##bc”, and “##z”. An input text “abcd” can be matched to a vocabulary token by walking from the root (upper left) and following the trie edges with labels “a”, “b”, “c”, “d” one by one. (The leading “##” symbols are special characters used in WordPiece tokenization that are described in more detail below.)

Trie diagram of the vocabulary [“a”, “abcd”, “##b”, “##bc”, “##z”]. Circles and arrows represent nodes and edges along the trie, respectively.

Second, inspired by the Aho-Corasick algorithm, a classical string-searching algorithm invented in 1975, we introduce a method that breaks out of a trie branch that fails to match the given input and skips directly to an alternative branch to continue matching. As in standard trie matching, during tokenization, we follow the trie edges to match the input characters one by one. When trie matching cannot match an input character for a given node, a standard algorithm would backtrack to the last character where a token was matched and then restart the trie matching procedure from there, which results in repetitive and wasteful iterations. Instead of backtracking, our method triggers a failure transition, which is done in two steps: (1) it collects the precomputed tokens stored at that node, which we call failure pops; and (2) it then follows the precomputed failure link to a new node from which the trie matching process continues.

For example, given a model with the vocabulary described above (“a”, “abcd”, “##b”, “##bc”, and “##z”), WordPiece tokenization distinguishes subword tokens matching at the start of the input word from the subword tokens starting in the middle (the latter being marked with two leading hashes “##”). Hence, for input text “abcz”, the expected tokenization output is [“a”, “##bc”, “##z”], where “a” matches at the beginning of the input while “##bc” and “##z” match in the middle. For this example, the figure below shows that, after successfully matching three characters ‘a’, ‘b’, ‘c’, trie matching cannot match the next character ‘z’ because “abcz” is not in the vocabulary. In this situation, LinMaxMatch conducts a failure transition by outputting the first recognized token (using the failure pop token “a”) and following the failure link to a new node to continue the matching process (in this case, node with “##bc” as the failure pop tokens).The process then repeats from the new node.

Trie structure for the same vocabulary as shown in the example above, now illustrating the approach taken by our new Fast WordPiece Tokenizer algorithm. Failure pops are bracketed and shown in purple. Failure links between nodes are indicated with dashed red line arrows.

Since at least n operations are required to read the entire input, the LinMaxMatch algorithm is asymptotically optimal for the MaxMatch problem.

End-to-End WordPiece Tokenization
Whereas the existing systems pre-tokenize the input text (splitting it into words by punctuation and whitespace characters) and then call WordPiece tokenization on each resulting word, we propose an end-to-end WordPiece tokenizer that combines pre-tokenization and WordPiece into a single, linear-time pass. It uses the LinMaxMatch trie matching and failure transitions as much as possible and only checks for punctuation and whitespace characters among the relatively few input characters that are not handled by the loop. It is more efficient as it traverses the input only once, performs fewer punctuation / whitespace checks, and skips the creation of intermediate words.

End-to-End WordPiece Tokenization.

Benchmark Results
We benchmark our method against two widely-adopted WordPiece tokenization implementations, HuggingFace Tokenizers, from the HuggingFace Transformer library, one of the most popular open-source NLP tools, and TensorFlow Text, the official library of text utilities for TensorFlow. We use the WordPiece vocabulary released with the BERT-Base, Multilingual Cased model.

We compared our algorithms with HuggingFace and TensorFlow Text on a large corpus (several million words) and found that the way the strings are split into tokens is identical to other implementations for both single-word and end-to-end tokenization.

To generate the test data, we sample 1,000 sentences from the multilingual Wikipedia dataset, covering 82 languages. On average, each word has four characters, and each sentence has 82 characters or 17 words. We found this dataset large enough because a much larger dataset (consisting of hundreds of thousands of sentences) generated similar results.

We compare the average runtime when tokenizing a single word or general text (end-to-end) for each system. Fast WordPiece tokenizer is 8.2x faster than HuggingFace and 5.1x faster than TensorFlow Text, on average, for general text end-to-end tokenization.

Average runtime of each system. Note that for better visualization, single-word tokenization and end-to-end tokenization are shown in different scales.

We also examine how the runtime grows with respect to the input length for single-word tokenization. Because of its linear-time complexity, the runtime of LinMaxMatch increases at most linearly with the input length, which is much slower than other quadratic-time approaches.

The average runtime of each system with respect to the input length for single-word tokenization.

We proposed LinMaxMatch for single-word WordPiece tokenization, which solves the decades-old MaxMatch problem in the asymptotically-optimal time with respect to the input length. LinMaxMatch extends the Aho-Corasick Algorithm, and the idea can be applied to more string search and transducer challenges. We also proposed an End-to-End WordPiece algorithm that combines pre-tokenization and WordPiece tokenization into a single, linear-time pass for even higher efficiency.

We gratefully acknowledge the key contributions and useful advices from other team members and colleagues, including Abbas Bazzi, Alexander Frömmgen, Alex Salcianu, Andrew Hilton, Bradley Green, Ed Chi, Chen Chen, Dave Dopson, Eric Lehman, Fangtao Li, Gabriel Schubiner, Gang Li, Greg Billock, Hong Wang, Jacob Devlin, Jayant Madhavan, JD Chen, Jifan Zhu, Jing Li, John Blitzer, Kirill Borozdin, Kristina Toutanova, Majid Hadian-Jazi, Mark Omernick, Max Gubin, Michael Fields, Michael Kwong, Namrata Godbole, Nathan Lintz, Pandu Nayak, Pew Putthividhya, Pranav Khaitan, Robby Neale, Ryan Doherty, Sameer Panwar, Sundeep Tirumalareddy, Terry Huang, Thomas Strohmann, Tim Herrmann, Tom Small, Tomer Shani, Wenwei Yu, Xiaoxue Zang, Xin Li, Yang Guo, Yang Song, Yiming Xiao, Yuan Shen, and many more.


More Efficient In-Context Learning with GLaM

Large language models (e.g., GPT-3) have many significant capabilities, such as performing few-shot learning across a wide array of tasks, including reading comprehension and question answering with very few or no training examples. While these models can perform better by simply using more parameters, training and serving these large models can be very computationally intensive. Is it possible to train and use these models more efficiently?

In pursuit of that question, today we introduce the Generalist Language Model (GLaM), a trillion weight model that can be trained and served efficiently (in terms of computation and energy use) thanks to sparsity, and achieves competitive performance on multiple few-shot learning tasks. GLaM’s performance compares favorably to a dense language model, GPT-3 (175B) with significantly improved learning efficiency across 29 public NLP benchmarks in seven categories, spanning language completion, open-domain question answering, and natural language inference tasks.

To build GLaM, we began by building a high-quality 1.6 trillion token dataset containing language usage representative of a wide range of downstream use-cases for the model. Web pages constitute the vast quantity of data in this unlabelled corpus, but their quality ranges from professional writing to low-quality comment and forum pages. We then developed a text quality filter that was trained on a collection of text from Wikipedia and books (both of which are generally higher quality sources) to determine the quality of the content for a webpage. Finally, we applied this filter to generate the final subset of webpages and combined this with books and Wikipedia to create the final training dataset.

Model and Architecture
GLaM is a mixture of experts (MoE) model, a type of model that can be thought of as having different submodels (or experts) that are each specialized for different inputs. The experts in each layer are controlled by a gating network that activates experts based on the input data. For each token (generally a word or part of a word), the gating network selects the two most appropriate experts to process the data. The full version of GLaM has 1.2T total parameters across 64 experts per MoE layer with 32 MoE layers in total, but only activates a subnetwork of 97B (8% of 1.2T) parameters per token prediction during inference.

The architecture of GLaM where each input token is dynamically routed to two selected expert networks out of 64 for prediction.

Similar to the GShard MoE Transformer, we replace the single feedforward network (the simplest layer of an artificial neural network, “Feedforward or FFN” in the blue boxes) of every other transformer layer with a MoE layer. This MoE layer has multiple experts, each a feedforward network with identical architecture but different weight parameters. Even though this MoE layer has many more parameters, the experts are sparsely activated, meaning that for a given input token, only two experts are used, giving the model more capacity while limiting computation. During training, each MoE layer’s gating network is trained to use its input to activate the best two experts for each token, which are then used for inference. For a MoE layer of E experts, this essentially provides a collection of E×(E-1) different feedforward network combinations (instead of one as in the classic Transformer architecture), leading to more computational flexibility.

The final learned representation of a token will be the weighted combination of the outputs from the two experts. This allows different experts to activate on different types of inputs. To enable scaling to larger models, each expert within the GLaM architecture can span multiple computational devices. We use the GSPMD compiler backend to solve the challenges in scaling the experts and train several variants (based on expert size and number of experts) of this architecture to understand the scaling effects of sparsely activated language models.

We use a zero-shot and one-shot setting where the tasks are never seen during training. The benchmarks for evaluation include (1) cloze and completion tasks [1,2,3]; (2) Open-domain question answering [4,5,6]; (3) Winograd-style tasks [7,8]; (4) commonsense reasoning [9,10,11]; (5) in-context reading comprehension [12,13,14,15,16]; (6) the SuperGLUE tasks; and (7) natural language inference [17]. In total, there are eight natural language generation tasks (NLG) where the generated phrases are evaluated against the ground truth targets via Exact Match (EM) accuracy and F1 measure, and 21 language understanding tasks (NLU) where the prediction from several options is chosen via conditional log-likelihood. Some tasks have variants and SuperGLUE consists of multiple tasks. Both EM accuracy and F1 are scaled from 0 to 100 across all our results and averaged for the NLG score below. The NLU score is an average of accuracy and F1 scores.

GLaM reduces to a basic dense Transformer-based language model architecture when each MoE layer only has one expert. In all experiments, we adopt the notation of (base dense model size) / (number of experts per MoE layer) to describe the GLaM model. For example, 1B/64E represents the architecture of a 1B parameter dense model with every other layer replaced by a 64 expert MoE layer. In the following sections, we explore GLaM’s performance and scaling properties, including baseline dense models trained on the same datasets. Compared with the recently announced Megatron-Turing model, GLaM is on-par on the seven respective tasks if using a 5% margin, while using 5x less computation during inference.

Below, we show the 1.2T-parameter sparsely activated model (GLaM) achieved higher results on average and on more tasks than the 175B-parameter dense GPT-3 model while using less computation during inference.

Average score for GLaM and GPT-3 on NLG (left) and NLU (right) tasks (higher is better).

Below we show a summary of the performance on 29 benchmarks compared to the dense model (GPT-3, 175B). GLaM exceeds or is on-par with the performance of the dense model on almost 80% of zero-shot tasks and almost 90% of one-shot tasks.

Evaluation Higher (>+5%) On-par (within 5%) Lower (<-5%)
Zero-shot 13 11 5
One-shot 14 10 5

Moreover, while the full version of GLaM has 1.2T total parameters, it only activates a subnetwork of 97B parameters (8% of 1.2T) per token during inference.

GLaM (64B/64E) GPT-3 (175B)
Total Parameters 1.162T 0.175T
Activated Parameters 0.097T 0.175T

Scaling Behavior
GLaM has two ways to scale: 1) scale the number of experts per layer, where each expert is hosted within one computation device, or 2) scale the size of each expert to go beyond the limit of a single device. To evaluate the scaling properties, we compare the respective dense model (FFN layers instead of MoE layers) of similar FLOPS per token at inference time.

Average zero-shot and one-shot performance by increasing the size of each expert. The FLOPS per token prediction at inference time increases as the expert size grows.

As shown above, performance across tasks scales with the size of the experts. GLaM sparsely activated models also perform better than dense models for similar FLOPs during inference for generation tasks. For understanding tasks, we observed that they perform similarly at smaller scales, but sparsely activated models outperform at larger scales.

Data Efficiency
Training large language models is computationally intensive, so efficiency improvements are useful to reduce energy consumption.

Below we show the computation costs for the full version of GLaM.

Computation cost in GFLOPS both for inference, per token (left) and for training (right).

These compute costs show that GLaM uses more computation during training since it trains on more tokens, but uses significantly less computation during inference. We show comparisons using different numbers of tokens to train below.

We also evaluated the learning curves of our models compared to the dense baseline.

Average zero-shot and one-shot performance of sparsely-activated and dense models on eight generative tasks as more tokens are processed in training.
Average zero-shot and one-shot performance of sparsely-activated and dense models on 21 understanding tasks as more tokens are processed in training.

The results above show that sparsely activated models need to train with significantly less data than dense models to reach similar zero-shot and one-shot performance, and if the same amount of data is used, sparsely activated models perform significantly better.

Finally, we assessed the energy efficiency of GLaM.

Comparison of power consumption during training.

While GLaM uses more computation during training, thanks to the more efficient software implementation powered by GSPMD and the advantage of TPUv4, it uses less power to train than other models.

Our large-scale sparsely activated language model, GLaM, achieves competitive results on zero-shot and one-shot learning and is a more efficient model than prior monolithic dense counterparts. We also show quantitatively that a high-quality dataset is essential for large language models. We hope that our work will spark more research into compute-efficient language models.

We wish to thank Claire Cui, Zhifeng Chen, Yonghui Wu, Quoc Le, Macduff Hughes, Fernando Pereira, Zoubin Ghahramani‎ and Jeff Dean for their support and invaluable input. Special thanks to our collaborators: Yanping Huang, Simon Tong, Yanqi Zhou, Yuanzhong Xu, Dmitry Lepikhin, Orhan Firat, Maxim Krikun, Tao Wang, Noam Shazeer, Barret Zoph, Liam Fedus, Maarten Bosma, Kun Zhang, Emma Wang, David Patterson, Zongwei Zhou, Naveen Kumar, Adams Yu, Laurent Shafey, Jonathan Shen, Ben Lee, Anmol Gulati, David So, Marie Pellat, Kevin Robinson, Kathy Meier-Hellstern‎, Aakanksha Chowdhery, Sharan Narang, Erica Moreira and Eric Ni for helpful discussions and inspirations; and the larger Google Research team. We would also like to thank Tom Small for the animated figure used in this post.