Category: Offsites

Posted by Junfeng He, Senior Research Scientist, and Kai Kohlhoff, Staff Research Scientist, Google Research

People have the remarkable ability to take in a tremendous amount of information (estimated to be ~10¹⁰ bits/s entering the retina) and selectively attend to a few task-relevant and interesting regions for further processing (e.g., memory, comprehension, action). Modeling human attention (the result of which is often called a saliency model) has therefore been of interest across the fields of neuroscience, psychology, human-computer interaction (HCI) and computer vision. The ability to predict which regions are likely to attract attention has numerous important applications in areas like graphics, photography, image compression and processing, and the measurement of visual quality.

We’ve previously discussed the possibility of accelerating eye movement research using machine learning and smartphone-based gaze estimation, which earlier required specialized hardware costing up to $30,000 per unit. Related research includes “Look to Speak”, which helps users with accessibility needs (e.g., people with ALS) to communicate with their eyes, and the recently published “Differentially private heatmaps” technique to compute heatmaps, like those for attention, while protecting users’ privacy.

In this blog, we present two papers (one from CVPR 2022, and one just accepted to CVPR 2023) that highlight our recent research in the area of human attention modeling: “Deep Saliency Prior for Reducing Visual Distraction” and “Learning from Unique Perspectives: User-aware Saliency Modeling”, together with recent research on saliency driven progressive loading for image compression (1, 2). We showcase how predictive models of human attention can enable delightful user experiences such as image editing to minimize visual clutter, distraction or artifacts, image compression for faster loading of webpages or apps, and guiding ML models towards more intuitive human-like interpretation and model performance. We focus on image editing and image compression, and discuss recent advances in modeling in the context of these applications.

Attention-guided image editing

Human attention models usually take an image as input (e.g., a natural image or a screenshot of a webpage), and predict a heatmap as output. The predicted heatmap on the image is evaluated against ground-truth attention data, which are typically collected by an eye tracker or approximated via mouse hovering/clicking. Previous models leveraged handcrafted features for visual clues, like color/brightness contrast, edges, and shape, while more recent approaches automatically learn discriminative features based on deep neural networks, from convolutional and recurrent neural networks to more recent vision transformer networks.

In “Deep Saliency Prior for Reducing Visual Distraction” (more information on this project site), we leverage deep saliency models for dramatic yet visually realistic edits, which can significantly change an observer’s attention to different image regions. For example, removing distracting objects in the background can reduce clutter in photos, leading to increased user satisfaction. Similarly, in video conferencing, reducing clutter in the background may increase focus on the main speaker (example demo here).

To explore what types of editing effects can be achieved and how these affect viewers’ attention, we developed an optimization framework for guiding visual attention in images using a differentiable, predictive saliency model. Our method employs a state-of-the-art deep saliency model. Given an input image and a binary mask representing the distractor regions, pixels within the mask will be edited under the guidance of the predictive saliency model such that the saliency within the masked region is reduced. To make sure the edited image is natural and realistic, we carefully choose four image editing operators: two standard image editing operations, namely recolorization and image warping (shift); and two learned operators (we do not define the editing operation explicitly), namely a multi-layer convolution filter, and a generative model (GAN).

With those operators, our framework can produce a variety of powerful effects, with examples in the figure below, including recoloring, inpainting, camouflage, object editing or insertion, and facial attribute editing. Importantly, all these effects are driven solely by the single, pre-trained saliency model, without any additional supervision or training. Note that our goal is not to compete with dedicated methods for producing each effect, but rather to demonstrate how multiple editing operations can be guided by the knowledge embedded within deep saliency models.

Examples of reducing visual distractions, guided by the saliency model with several operators. The distractor region is marked on top of the saliency map (red border) in each example.

Enriching experiences with user-aware saliency modeling

Prior research assumes a single saliency model for the whole population. However, human attention varies between individuals — while the detection of salient clues is fairly consistent, their order, interpretation, and gaze distributions can differ substantially. This offers opportunities to create personalized user experiences for individuals or groups. In “Learning from Unique Perspectives: User-aware Saliency Modeling”, we introduce a user-aware saliency model, the first that can predict attention for one user, a group of users, and the general population, with a single model.

As shown in the figure below, core to the model is the combination of each participant’s visual preferences with a per-user attention map and adaptive user masks. This requires per-user attention annotations to be available in the training data, e.g., the OSIE mobile gaze dataset for natural images; FiWI and WebSaliency datasets for web pages. Instead of predicting a single saliency map representing attention of all users, this model predicts per-user attention maps to encode individuals’ attention patterns. Further, the model adopts a user mask (a binary vector with the size equal to the number of participants) to indicate the presence of participants in the current sample, which makes it possible to select a group of participants and combine their preferences into a single heatmap.

An overview of the user aware saliency model framework. The example image is from OSIE image set.

During inference, the user mask allows making predictions for any combination of participants. In the following figure, the first two rows are attention predictions for two different groups of participants (with three people in each group) on an image. A conventional attention prediction model will predict identical attention heatmaps. Our model can distinguish the two groups (e.g., the second group pays less attention to the face and more attention to the food than the first). Similarly, the last two rows are predictions on a webpage for two distinctive participants, with our model showing different preferences (e.g., the second participant pays more attention to the left region than the first).

Predicted attention vs. ground truth (GT). EML-Net: predictions from a state-of-the-art model, which will have the same predictions for the two participants/groups. Ours: predictions from our proposed user aware saliency model, which can predict the unique preference of each participant/group correctly. The first image is from OSIE image set, and the second is from FiWI.

Progressive image decoding centered on salient features

Besides image editing, human attention models can also improve users’ browsing experience. One of the most frustrating and annoying user experiences while browsing is waiting for web pages with images to load, especially in conditions with low network connectivity. One way to improve the user experience in such cases is with progressive decoding of images, which decodes and displays increasingly higher-resolution image sections as data are downloaded, until the full-resolution image is ready. Progressive decoding usually proceeds in a sequential order (e.g., left to right, top to bottom). With a predictive attention model (1, 2), we can instead decode images based on saliency, making it possible to send the data necessary to display details of the most salient regions first. For example, in a portrait, bytes for the face can be prioritized over those for the out-of-focus background. Consequently, users perceive better image quality earlier and experience significantly reduced wait times. More details can be found in our open source blog posts (post 1, post 2). Thus, predictive attention models can help with image compression and faster loading of web pages with images, improve rendering for large images and streaming/VR applications.

Conclusion

We’ve shown how predictive models of human attention can enable delightful user experiences via applications such as image editing that can reduce clutter, distractions or artifacts in images or photos for users, and progressive image decoding that can greatly reduce the perceived waiting time for users while images are fully rendered. Our user-aware saliency model can further personalize the above applications for individual users or groups, enabling richer and more unique experiences.

Another interesting direction for predictive attention models is whether they can help improve robustness of computer vision models in tasks such as object classification or detection. For example, in “Teacher-generated spatial-attention labels boost robustness and accuracy of contrastive models”, we show that a predictive human attention model can guide contrastive learning models to achieve better representation and improve the accuracy/robustness of classification tasks (on the ImageNet and ImageNet-C datasets). Further research in this direction could enable applications such as using radiologist’s attention on medical images to improve health screening or diagnosis, or using human attention in complex driving scenarios to guide autonomous driving systems.

Acknowledgements

This work involved collaborative efforts from a multidisciplinary team of software engineers, researchers, and cross-functional contributors. We’d like to thank all the co-authors of the papers/research, including Kfir Aberman, Gamaleldin F. Elsayed, Moritz Firsching, Shi Chen, Nachiappan Valliappan, Yushi Yao, Chang Ye, Yossi Gandelsman, Inbar Mosseri, David E. Jacobes, Yael Pritch, Shaolei Shen, and Xinyu Ye. We also want to thank team members Oscar Ramirez, Venky Ramachandran and Tim Fujita for their help. Finally, we thank Vidhya Navalpakkam for her technical leadership in initiating and overseeing this body of work.

Posted by Su Wang and Ceslee Montgormery, Research Engineers, Google Research

In the last few years, text-to-image generation research has seen an explosion of breakthroughs (notably, Imagen, Parti, DALL-E 2, etc.) that have naturally permeated into related topics. In particular, text-guided image editing (TGIE) is a practical task that involves editing generated and photographed visuals rather than completely redoing them. Quick, automated, and controllable editing is a convenient solution when recreating visuals would be time-consuming or infeasible (e.g., tweaking objects in vacation photos or perfecting fine-grained details on a cute pup generated from scratch). Further, TGIE represents a substantial opportunity to improve training of foundational models themselves. Multimodal models require diverse data to train properly, and TGIE editing can enable the generation and recombination of high-quality and scalable synthetic data that, perhaps most importantly, can provide methods to optimize the distribution of training data along any given axis.

In “Imagen Editor and EditBench: Advancing and Evaluating Text-Guided Image Inpainting”, to be presented at CVPR 2023, we introduce Imagen Editor, a state-of-the-art solution for the task of masked inpainting — i.e., when a user provides text instructions alongside an overlay or “mask” (usually generated within a drawing-type interface) indicating the area of the image they would like to modify. We also introduce EditBench, a method that gauges the quality of image editing models. EditBench goes beyond the commonly used coarse-grained “does this image match this text” methods, and drills down to various types of attributes, objects, and scenes for a more fine-grained understanding of model performance. In particular, it puts strong emphasis on the faithfulness of image-text alignment without losing sight of image quality.

Given an image, a user-defined mask, and a text prompt, Imagen Editor makes localized edits to the designated areas. The model meaningfully incorporates the user’s intent and performs photorealistic edits.

Imagen Editor

Imagen Editor is a diffusion-based model fine-tuned on Imagen for editing. It targets improved representations of linguistic inputs, fine-grained control and high-fidelity outputs. Imagen Editor takes three inputs from the user: 1) the image to be edited, 2) a binary mask to specify the edit region, and 3) a text prompt — all three inputs guide the output samples.

Imagen Editor depends on three core techniques for high-quality text-guided image inpainting. First, unlike prior inpainting models (e.g., Palette, Context Attention, Gated Convolution) that apply random box and stroke masks, Imagen Editor employs an object detector masking policy with an object detector module that produces object masks during training. Object masks are based on detected objects rather than random patches and allow for more principled alignment between edit text prompts and masked regions. Empirically, the method helps the model stave off the prevalent issue of the text prompt being ignored when masked regions are small or only partially cover an object (e.g., CogView2).

Random masks (left) frequently capture background or intersect object boundaries, defining regions that can be plausibly inpainted just from image context alone. Object masks (right) are harder to inpaint from image context alone, encouraging models to rely more on text inputs during training.

Next, during training and inference, Imagen Editor enhances high resolution editing by conditioning on full resolution (1024×1024 in this work), channel-wise concatenation of the input image and the mask (similar to SR3, Palette, and GLIDE). For the base diffusion 64×64 model and the 64×64→256×256 super-resolution models, we apply a parameterized downsampling convolution (e.g., convolution with a stride), which we empirically find to be critical for high fidelity.

Imagen is fine-tuned for image editing. All of the diffusion models, i.e., the base model and super-resolution (SR) models, are conditioned on high-resolution 1024×1024 image and mask inputs. To this end, new convolutional image encoders are introduced.

Finally, at inference we apply classifier-free guidance (CFG) to bias samples to a particular conditioning, in this case, text prompts. CFG interpolates between the text-conditioned and unconditioned model predictions to ensure strong alignment between the generated image and the input text prompt for text-guided image inpainting. We follow Imagen Video and use high guidance weights with guidance oscillation (a guidance schedule that oscillates within a value range of guidance weights). In the base model (the stage-1 64x diffusion), where ensuring strong alignment with text is most critical, we use a guidance weight schedule that oscillates between 1 and 30. We observe that high guidance weights combined with oscillating guidance result in the best trade-off between sample fidelity and text-image alignment.

EditBench

The EditBench dataset for text-guided image inpainting evaluation contains 240 images, with 120 generated and 120 natural images. Generated images are synthesized by Parti and natural images are drawn from the Visual Genome and Open Images datasets. EditBench captures a wide variety of language, image types, and levels of text prompt specificity (i.e., simple, rich, and full captions). Each example consists of (1) a masked input image, (2) an input text prompt, and (3) a high-quality output image used as reference for automatic metrics. To provide insight into the relative strengths and weaknesses of different models, EditBench prompts are designed to test fine-grained details along three categories: (1) attributes (e.g., material, color, shape, size, count); (2) object types (e.g., common, rare, text rendering); and (3) scenes (e.g., indoor, outdoor, realistic, or paintings). To understand how different specifications of prompts affect model performance, we provide three text prompt types: a single-attribute (Mask Simple) or a multi-attribute description of the masked object (Mask Rich) – or an entire image description (Full Image). Mask Rich, especially, probes the models’ ability to handle complex attribute binding and inclusion.

The full image is used as a reference for successful inpainting. The mask covers the target object with a free-form, non-hinting shape. We evaluate Mask Simple, Mask Rich and Full Image prompts, consistent with conventional text-to-image models.

Due to the intrinsic weaknesses in existing automatic evaluation metrics (CLIPScore and CLIP-R-Precision) for TGIE, we hold human evaluation as the gold standard for EditBench. In the section below, we demonstrate how EditBench is applied to model evaluation.

Evaluation

We evaluate the Imagen Editor model — with object masking (IM) and with random masking (IM-RM) — against comparable models, Stable Diffusion (SD) and DALL-E 2 (DL2). Imagen Editor outperforms these models by substantial margins across all EditBench evaluation categories.

For Full Image prompts, single-image human evaluation provides binary answers to confirm if the image matches the caption. For Mask Simple prompts, single-image human evaluation confirms if the object and attribute are properly rendered, and bound correctly (e.g., for a red cat, a white cat on a red table would be an incorrect binding). Side-by-side human evaluation uses Mask Rich prompts only for side-by-side comparisons between IM and each of the other three models (IM-RM, DL2, and SD), and indicates which image matches with the caption better for text-image alignment, and which image is most realistic.

Human evaluation. Full Image prompts elicit annotators’ overall impression of text-image alignment; Mask Simple and Mask Rich check for the correct inclusion of particular attributes, objects and attribute binding.

For single-image human evaluation, IM receives the highest ratings across-the-board (10–13% higher than the 2nd-highest performing model). For the rest, the performance order is IM-RM > DL2 > SD (with 3–6% difference) except for with Mask Simple, where IM-RM falls 4-8% behind. As relatively more semantic content is involved in Full and Mask Rich, we conjecture IM-RM and IM are benefited by the higher performing T5 XXL text encoder.

Single-image human evaluations of text-guided image inpainting on EditBench by prompt type. For Mask Simple and Mask Rich prompts, text-image alignment is correct if the edited image accurately includes every attribute and object specified in the prompt, including the correct attribute binding. Note that due to different evaluation designs, Full vs. Mask-only prompts, results are less directly comparable.

EditBench focuses on fine-grained annotation, so we evaluate models for object and attribute types. For object types, IM leads in all categories, performing 10–11% better than the 2nd-highest performing model in common, rare, and text-rendering.

Single-image human evaluations on EditBench Mask Simple by object type. As a cohort, models are better at object rendering than text-rendering.

For attribute types, IM is rated much higher (13–16%) than the 2nd highest performing model, except for in count, where DL2 is merely 1% behind.

Single-image human evaluations on EditBench Mask Simple by attribute type. Object masking improves adherence to prompt attributes across-the-board (IM vs. IM-RM).

Side-by-side compared with other models one-vs-one, IM leads in text alignment with a substantial margin, being preferred by annotators compared to SD, DL2, and IM-RM.

Side-by-side human evaluation of image realism & text-image alignment on EditBench Mask Rich prompts. For text-image alignment, Imagen Editor is preferred in all comparisons.

Finally, we illustrate a representative side-by-side comparative for all the models. See the paper for more examples.

Example model outputs for Mask Simple vs. Mask Rich prompts. Object masking improves Imagen Editor’s fine-grained adherence to the prompt compared to the same model trained with random masking.

Conclusion

We presented Imagen Editor and EditBench, making significant advancements in text-guided image inpainting and the evaluation thereof. Imagen Editor is a text-guided image inpainting fine-tuned from Imagen. EditBench is a comprehensive systematic benchmark for text-guided image inpainting, evaluating performance across multiple dimensions: attributes, objects, and scenes. Note that due to concerns in relation to responsible AI, we are not releasing Imagen Editor to the public. EditBench on the other hand is released in full for the benefit of the research community.

Acknowledgments

Thanks to Gunjan Baid, Nicole Brichtova, Sara Mahdavi, Kathy Meier-Hellstern, Zarana Parekh, Anusha Ramesh, Tris Warkentin, Austin Waters, and Vijay Vasudevan for their generous support. We give thanks to Igor Karpov, Isabel Kraus-Liang, Raghava Ram Pamidigantam, Mahesh Maddinala, and all the anonymous human annotators for their coordination to complete the human evaluation tasks. We are grateful to Huiwen Chang, Austin Tarango, and Douglas Eck for providing paper feedback. Thanks to Erica Moreira and Victor Gomes for help with resource coordination. Finally, thanks to the authors of DALL-E 2 for giving us permission to use their model outputs for research purposes.

Posted by Thibault Sellam, Research Scientist, Google

Previously, we presented the 1,000 languages initiative and the Universal Speech Model with the goal of making speech and language technologies available to billions of users around the world. Part of this commitment involves developing high-quality speech synthesis technologies, which build upon projects such as VDTTS and AudioLM, for users that speak many different languages.

After developing a new model, one must evaluate whether the speech it generates is accurate and natural: the content must be relevant to the task, the pronunciation correct, the tone appropriate, and there should be no acoustic artifacts such as cracks or signal-correlated noise. Such evaluation is a major bottleneck in the development of multilingual speech systems.

The most popular method to evaluate the quality of speech synthesis models is human evaluation: a text-to-speech (TTS) engineer produces a few thousand utterances from the latest model, sends them for human evaluation, and receives results a few days later. This evaluation phase typically involves listening tests, during which dozens of annotators listen to the utterances one after the other to determine how natural they sound. While humans are still unbeaten at detecting whether a piece of text sounds natural, this process can be impractical — especially in the early stages of research projects, when engineers need rapid feedback to test and restrategize their approach. Human evaluation is expensive, time consuming, and may be limited by the availability of raters for the languages of interest.

Another barrier to progress is that different projects and institutions typically use various ratings, platforms and protocols, which makes apples-to-apples comparisons impossible. In this regard, speech synthesis technologies lag behind text generation, where researchers have long complemented human evaluation with automatic metrics such as BLEU or, more recently, BLEURT.

In “SQuId: Measuring Speech Naturalness in Many Languages“, to be presented at ICASSP 2023, we introduce SQuId (Speech Quality Identification), a 600M parameter regression model that describes to what extent a piece of speech sounds natural. SQuId is based on mSLAM (a pre-trained speech-text model developed by Google), fine-tuned on over a million quality ratings across 42 languages and tested in 65. We demonstrate how SQuId can be used to complement human ratings for evaluation of many languages. This is the largest published effort of this type to date.

Evaluating TTS with SQuId

The main hypothesis behind SQuId is that training a regression model on previously collected ratings can provide us with a low-cost method for assessing the quality of a TTS model. The model can therefore be a valuable addition to a TTS researcher’s evaluation toolbox, providing a near-instant, albeit less accurate alternative to human evaluation.

SQuId takes an utterance as input and an optional locale tag (i.e., a localized variant of a language, such as “Brazilian Portuguese” or “British English”). It returns a score between 1 and 5 that indicates how natural the waveform sounds, with a higher value indicating a more natural waveform.

Internally, the model includes three components: (1) an encoder, (2) a pooling / regression layer, and (3) a fully connected layer. First, the encoder takes a spectrogram as input and embeds it into a smaller 2D matrix that contains 3,200 vectors of size 1,024, where each vector encodes a time step. The pooling / regression layer aggregates the vectors, appends the locale tag, and feeds the result into a fully connected layer that returns a score. Finally, we apply application-specific post-processing that rescales or normalizes the score so it is within the [1, 5] range, which is common for naturalness human ratings. We train the whole model end-to-end with a regression loss.

The encoder is by far the largest and most important piece of the model. We used mSLAM, a pre-existing 600M-parameter Conformer pre-trained on both speech (51 languages) and text (101 languages).

The SQuId model.

To train and evaluate the model, we created the SQuId corpus: a collection of 1.9 million rated utterances across 66 languages, collected for over 2,000 research and product TTS projects. The SQuId corpus covers a diverse array of systems, including concatenative and neural models, for a broad range of use cases, such as driving directions and virtual assistants. Manual inspection reveals that SQuId is exposed to a vast range of of TTS errors, such as acoustic artifacts (e.g., cracks and pops), incorrect prosody (e.g., questions without rising intonations in English), text normalization errors (e.g., verbalizing “7/7” as “seven divided by seven” rather than “July seventh”), or pronunciation mistakes (e.g., verbalizing “tough” as “toe”).

A common issue that arises when training multilingual systems is that the training data may not be uniformly available for all the languages of interest. SQuId was no exception. The following figure illustrates the size of the corpus for each locale. We see that the distribution is largely dominated by US English.

Locale distribution in the SQuId dataset.

How can we provide good performance for all languages when there are such variations? Inspired by previous work on machine translation, as well as past work from the speech literature, we decided to train one model for all languages, rather than using separate models for each language. The hypothesis is that if the model is large enough, then cross-locale transfer can occur: the model’s accuracy on each locale improves as a result of jointly training on the others. As our experiments show, cross-locale proves to be a powerful driver of performance.

Experimental results

To understand SQuId’s overall performance, we compare it to a custom Big-SSL-MOS model (described in the paper), a competitive baseline inspired by MOS-SSL, a state-of-the-art TTS evaluation system. Big-SSL-MOS is based on w2v-BERT and was trained on the VoiceMOS’22 Challenge dataset, the most popular dataset at the time of evaluation. We experimented with several variants of the model, and found that SQuId is up to 50.0% more accurate.

SQuId versus state-of-the-art baselines. We measure agreement with human ratings using the Kendall Tau, where a higher value represents better accuracy.

To understand the impact of cross-locale transfer, we run a series of ablation studies. We vary the amount of locales introduced in the training set and measure the effect on SQuId’s accuracy. In English, which is already over-represented in the dataset, the effect of adding locales is negligible.

SQuId’s performance on US English, using 1, 8, and 42 locales during fine-tuning.

However, cross-locale transfer is much more effective for most other locales:

SQuId’s performance on four selected locales (Korean, French, Thai, and Tamil), using 1, 8, and 42 locales during fine-tuning. For each locale, we also provide the training set size.

To push transfer to its limit, we held 24 locales out during training and used them for testing exclusively. Thus, we measure to what extent SQuId can deal with languages that it has never seen before. The plot below shows that although the effect is not uniform, cross-locale transfer works.

SQuId’s performance on four “zero-shot” locales; using 1, 8, and 42 locales during fine-tuning.

When does cross-locale operate, and how? We present many more ablations in the paper, and show that while language similarity plays a role (e.g., training on Brazilian Portuguese helps European Portuguese) it is surprisingly far from being the only factor that matters.

Conclusion and future work

We introduce SQuId, a 600M parameter regression model that leverages the SQuId dataset and cross-locale learning to evaluate speech quality and describe how natural it sounds. We demonstrate that SQuId can complement human raters in the evaluation of many languages. Future work includes accuracy improvements, expanding the range of languages covered, and tackling new error types.

Acknowledgements

The author of this post is now part of Google DeepMind. Many thanks to all authors of the paper: Ankur Bapna, Joshua Camp, Diana Mackinnon, Ankur P. Parikh, and Jason Riesa.

Posted by Ruofei Du, Research Scientist, and Alex Olwal, Senior Staff Research Scientist, Google Augmented Reality

Recent advances in video conferencing have significantly improved remote video communication through features like live captioning and noise cancellation. However, there are various situations where dynamic visual augmentation would be useful to better convey complex and nuanced information. For example, when discussing what to order at a Japanese restaurant, your friends could share visuals that would help you feel more confident about ordering the “Sukiyaki”. Or when talking about your recent family trip to San Francisco, you may want to show a photo from your personal album.

In “Visual Captions: Augmenting Verbal Communication With On-the-fly Visuals”, presented at ACM CHI 2023, we introduce a system that uses verbal cues to augment synchronous video communication with real-time visuals. We fine-tuned a large language model to proactively suggest relevant visuals in open-vocabulary conversations using a dataset we curated for this purpose. We open sourced Visual Captions as part of the ARChat project, which is designed for rapid prototyping of augmented communication with real-time transcription.

Visual Captions facilitates verbal communication with real-time visuals. The system is even robust against typical mistakes that may often appear in real-time speech-to-text transcription. For example, out of context, the transcription model misunderstood the word “pier” as “pair”, but Visual Captions still recommends images of the Santa Monica Pier.

Design space for augmenting verbal communication with dynamic visuals

We invited 10 internal participants, each with various technical and non-technical backgrounds, including software engineers, researchers, UX designers, visual artists, students, etc., to discuss their particular needs and desires for a potential real-time visual augmentation service. In two sessions, we introduced low-fidelity prototypes of the envisioned system, followed by video demos of the existing text-to-image systems. These discussions informed a design space with eight dimensions for visual augmentation of real-time conversations, labeled below as D1 to D8.

Visual augmentations could be synchronous or asynchronous with the conversation (D1: Temporal), could be used for both expressing and understanding speech content (D2: Subject), and could be applied using a wide range of different visual content, visual types, and visual sources (D3: Visual). Such visual augmentation might vary depending on the scale of the meetings (D4: Scale) and whether a meeting is in co-located or remote settings (D5: Space). These factors also influence whether the visuals should be displayed privately, shared between participants, or public to everyone (D6: Privacy). Participants also identified different ways in which they would like to interact with the system while having conversations (D7: Initiation). For example, people proposed different levels of “proactivity”, which indicates the degree to which users would like the model to take the initiative. Finally, participants envisioned different methods of interaction, for example, using speech or gestures for input. (D8: Interaction).

Design space for augmenting verbal communication with dynamic visuals.

Informed by this initial feedback, we designed Visual Captions to focus on generating synchronous visuals of semantically relevant visual content, type, and source. While participants in these initial exploratory sessions were participating in one-to-one remote conversations, deployment of Visual Captions in the wild will often be in one-to-many (e.g., an individual giving a presentation to an audience) and many-to-many scenarios (e.g., a discussion among multiple people in a meeting).

Because the visual that best complements a conversation depends strongly on the context of the discussion, we needed a training set specific to this purpose. So, we collected a dataset of 1595 quadruples of language (1), visual content (2), type (3), and source (4) across a variety of contexts, including daily conversations, lectures, and travel guides. For example, “I would love to see it!” corresponds to visual content of “face smiling”, a visual type of “emoji”, and visual source of “public search”. “Did she tell you about our trip to Mexico?” corresponds to visual content of “a photo from the trip to Mexico”, a visual type of “photo”, and visual source of “personal album”. We publicly released this VC1.5K dataset for the research community.

Visual intent prediction model

To predict what visuals could supplement a conversation, we trained a visual intent prediction model based on a large language model using the VC1.5K dataset. For training, we parsed each visual intent into the format of “<Visual Type> of <Visual Content> from <Visual Source>“.

{"prompt": "<Previous Two Sentences> →", 
  "completion": 
"<Visual Type 1> of "<Visual Type 1> from "<Visual Source 1>;
 <Visual Type 2> of "<Visual Type 2> from "<Visual Source 2>; 
  ... 𝑛"}

Using this format, this system can handle open-vocabulary conversations and contextually predict visual content, visual source, and visual type. Anecdotally, we found that it outperforms keyword-based approaches, which fail to handle open-vocabulary examples like “Your aunt Amy will be visiting this Saturday,” and cannot suggest relevant visual types or visual sources.

Examples of visual intent predictions by our model.

We used 1276 (80%) examples from the VC1.5K dataset for fine-tuning the large language model and the remaining 319 (20%) examples as test data. We measured the performance of the fine-tuned model with the token accuracy metric, i.e., the percentage of tokens in a batch that were correctly predicted by the model. During training, our model reached a training token accuracy of 97% and a validation token accuracy of 87%.

Performance

To evaluate the utility of the trained Visual Captions model, we invited 89 participants to perform 846 tasks. They were asked to provide feedback on a scale of “1 — Strongly Disagree” to “7 — Strongly Agree” for six qualitative statements. Most participants preferred to have the visual during a conversation (Q1, 83% ≥ 5–Somewhat Agree). Moreover, they considered the displayed visuals to be useful and informative (Q2, 82% ≥ 5–Somewhat Agree), high-quality (Q3, 82% ≥ 5–Somewhat Agree), and relevant to the original speech (Q4, 84% ≥ 5–Somewhat Agree). Participants also found the predicted visual type (Q5, 87% ≥ 5–Somewhat Agree) and visual source (Q6, 86% ≥ 5–Somewhat Agree) to be accurate given the context of the corresponding conversation.

Technical evaluation results of the visual prediction model rated by study participants.

With this fine-tuned visual intent prediction model, we developed Visual Captions on the ARChat platform, which can add new interactive widgets directly on the camera streams of video conferencing platforms, such as Google Meet. As shown in the system workflow below, Visual Captions automatically captures the user’s speech, retrieves the last sentences, feeds them into the visual intent prediction model every 100 ms, retrieves relevant visuals, and then suggests visuals in real time.

System workflow of Visual Captions.

Visual Captions provides three levels of proactivity when suggesting visuals:

• Auto-display (high-proactivity): The system autonomously searches and displays visuals publicly to all meeting participants. No user interaction required.
• Auto-suggest (medium-proactivity): The suggested visuals are shown in a private scrolling view. A user then clicks a visual to display it publicly. In this mode, the system is proactively recommending visuals, but the user decides when and what to display.
• On-demand-suggest (low-proactivity): The system will only suggest visuals if a user presses the spacebar.

Quantitative and qualitative evaluation: User studies

We evaluated Visual Captions in both a controlled lab study (n = 26) and in-the-wild deployment studies (n = 10). Participants found that real-time visuals facilitated live conversations by helping explain unfamiliar concepts, resolve language ambiguities, and make conversations more engaging. Participants also reported different preferences for interacting with the system in-situ, and that varying levels of proactivity were preferred in different social scenarios.

Participants’ Task Load Index and Likert scale ratings (from 1 – Strongly Disagree to 7 – Strongly Agree) of four conversations without Visual Captions (“No VC”) and the three Visual Captions modes: auto-display, auto-suggest, and on-demand suggest.

Conclusions and future directions

This work proposes a system for real-time visual augmentation of verbal communication, called Visual Captions, that was trained using a dataset of 1595 visual intents collected from 246 participants, covering 15 topic categories. We publicly release the training dataset, VC1.5K to the research community to support further research in this space. We have also deployed Visual Captions in ARChat, which facilitates video conferences in Google Meet by transcribing meetings and augmenting the camera video streams.

Visual Captions represents a significant step towards enhancing verbal communication with on-the-fly visuals. By understanding the importance of visual cues in everyday conversations, we can create more effective communication tools and improve how people connect.

Acknowledgements

This work is a collaboration across multiple teams at Google. Key contributors to the project include Xingyu “Bruce” Liu, Vladimir Kirilyuk, Xiuxiu Yuan, Peggy Chi, Alex Olwal, and Ruofei Du.

We would like to extend our thanks to those on the ARChat team who provided assistance, including Jason Mayes, Max Spear, Na Li, Jun Zhang, Jing Jin, Yuan Ren, Adarsh Kowdle, Ping Yu, Darcy Philippon, and Ezgi Oztelcan. We would also like to thank the many people with whom we’ve had insightful discussions and those who provided feedback on the manuscript, including Eric Turner, Yinda Zhang, Feitong Tan, Danhang Tang, and Shahram Izadi. We would also like to thank our CHI reviewers for their insightful feedback.

Posted by Arsha Nagrani and Paul Hongsuck Seo, Research Scientists, Google Research

Automatic speech recognition (ASR) is a well-established technology that is widely adopted for various applications such as conference calls, streamed video transcription and voice commands. While the challenges for this technology are centered around noisy audio inputs, the visual stream in multimodal videos (e.g., TV, online edited videos) can provide strong cues for improving the robustness of ASR systems — this is called audiovisual ASR (AV-ASR).

Although lip motion can provide strong signals for speech recognition and is the most common area of focus for AV-ASR, the mouth is often not directly visible in videos in the wild (e.g., due to egocentric viewpoints, face coverings, and low resolution) and therefore, a new emerging area of research is unconstrained AV-ASR (e.g., AVATAR), which investigates the contribution of entire visual frames, and not just the mouth region.

Building audiovisual datasets for training AV-ASR models, however, is challenging. Datasets such as How2 and VisSpeech have been created from instructional videos online, but they are small in size. In contrast, the models themselves are typically large and consist of both visual and audio encoders, and so they tend to overfit on these small datasets. Nonetheless, there have been a number of recently released large-scale audio-only models that are heavily optimized via large-scale training on massive audio-only data obtained from audio books, such as LibriLight and LibriSpeech. These models contain billions of parameters, are readily available, and show strong generalization across domains.

With the above challenges in mind, in “AVFormer: Injecting Vision into Frozen Speech Models for Zero-Shot AV-ASR”, we present a simple method for augmenting existing large-scale audio-only models with visual information, at the same time performing lightweight domain adaptation. AVFormer injects visual embeddings into a frozen ASR model (similar to how Flamingo injects visual information into large language models for vision-text tasks) using lightweight trainable adaptors that can be trained on a small amount of weakly labeled video data with minimum additional training time and parameters. We also introduce a simple curriculum scheme during training, which we show is crucial to enable the model to jointly process audio and visual information effectively. The resulting AVFormer model achieves state-of-the-art zero-shot performance on three different AV-ASR benchmarks (How2, VisSpeech and Ego4D), while also crucially preserving decent performance on traditional audio-only speech recognition benchmarks (i.e., LibriSpeech).

Unconstrained audiovisual speech recognition. We inject vision into a frozen speech model (BEST-RQ, in grey) for zero-shot audiovisual ASR via lightweight modules to create a parameter- and data-efficient model called AVFormer (blue). The visual context can provide helpful clues for robust speech recognition especially when the audio signal is noisy (the visual loaf of bread helps correct the audio-only mistake “clove” to “loaf” in the generated transcript).

Injecting vision using lightweight modules

Our goal is to add visual understanding capabilities to an existing audio-only ASR model while maintaining its generalization performance to various domains (both AV and audio-only domains).

To achieve this, we augment an existing state-of-the-art ASR model (Best-RQ) with the following two components: (i) linear visual projector and (ii) lightweight adapters. The former projects visual features in the audio token embedding space. This process allows the model to properly connect separately pre-trained visual feature and audio input token representations. The latter then minimally modifies the model to add understanding of multimodal inputs from videos. We then train these additional modules on unlabeled web videos from the HowTo100M dataset, along with the outputs of an ASR model as pseudo ground truth, while keeping the rest of the Best-RQ model frozen. Such lightweight modules enable data-efficiency and strong generalization of performance.

We evaluated our extended model on AV-ASR benchmarks in a zero-shot setting, where the model is never trained on a manually annotated AV-ASR dataset.

Curriculum learning for vision injection

After the initial evaluation, we discovered empirically that with a naïve single round of joint training, the model struggles to learn both the adapters and the visual projectors in one go. To mitigate this issue, we introduced a two-phase curriculum learning strategy that decouples these two factors — domain adaptation and visual feature integration — and trains the network in a sequential manner. In the first phase, the adapter parameters are optimized without feeding visual tokens at all. Once the adapters are trained, we add the visual tokens and train the visual projection layers alone in the second phase while the trained adapters are kept frozen.

The first stage focuses on audio domain adaptation. By the second phase, the adapters are completely frozen and the visual projector must simply learn to generate visual prompts that project the visual tokens into the audio space. In this way, our curriculum learning strategy allows the model to incorporate visual inputs as well as adapt to new audio domains in AV-ASR benchmarks. We apply each phase just once, as an iterative application of alternating phases leads to performance degradation.

Overall architecture and training procedure for AVFormer. The architecture consists of a frozen Conformer encoder-decoder model, and a frozen CLIP encoder (frozen layers shown in gray with a lock symbol), in conjunction with two lightweight trainable modules – (i) visual projection layer (orange) and bottleneck adapters (blue) to enable multimodal domain adaptation. We propose a two-phase curriculum learning strategy: the adapters (blue) are first trained without any visual tokens, after which the visual projection layer (orange) is tuned while all the other parts are kept frozen.

The plots below show that without curriculum learning, our AV-ASR model is worse than the audio-only baseline across all datasets, with the gap increasing as more visual tokens are added. In contrast, when the proposed two-phase curriculum is applied, our AV-ASR model performs significantly better than the baseline audio-only model.

Effects of curriculum learning. Red and blue lines are for audiovisual models and are shown on 3 datasets in the zero-shot setting (lower WER % is better). Using the curriculum helps on all 3 datasets (for How2 (a) and Ego4D (c) it is crucial for outperforming audio-only performance). Performance improves up until 4 visual tokens, at which point it saturates.

Results in zero-shot AV-ASR

We compare AVFormer to BEST-RQ, the audio version of our model, and AVATAR, the state of the art in AV-ASR, for zero-shot performance on the three AV-ASR benchmarks: How2, VisSpeech and Ego4D. AVFormer outperforms AVATAR and BEST-RQ on all, even outperforming both AVATAR and BEST-RQ when they are trained on LibriSpeech and the full set of HowTo100M. This is notable because for BEST-RQ, this involves training 600M parameters, while AVFormer only trains 4M parameters and therefore requires only a small fraction of the training dataset (5% of HowTo100M). Moreover, we also evaluate performance on LibriSpeech, which is audio-only, and AVFormer outperforms both baselines.

Comparison to state-of-the-art methods for zero-shot performance across different AV-ASR datasets. We also show performances on LibriSpeech which is audio-only. Results are reported as WER % (lower is better). AVATAR and BEST-RQ are finetuned end-to-end (all parameters) on HowTo100M whereas AVFormer works effectively even with 5% of the dataset thanks to the small set of finetuned parameters.

Conclusion

We introduce AVFormer, a lightweight method for adapting existing, frozen state-of-the-art ASR models for AV-ASR. Our approach is practical and efficient, and achieves impressive zero-shot performance. As ASR models get larger and larger, tuning the entire parameter set of pre-trained models becomes impractical (even more so for different domains). Our method seamlessly allows both domain transfer and visual input mixing in the same, parameter efficient model.

Acknowledgements

This research was conducted by Paul Hongsuck Seo, Arsha Nagrani and Cordelia Schmid.

Posted by Ziniu Hu, Student Researcher, and Alireza Fathi, Research Scientist, Google Research, Perception Team

Large-scale models, such as T5, GPT-3, PaLM, Flamingo and PaLI, have demonstrated the ability to store substantial amounts of knowledge when scaled to tens of billions of parameters and trained on large text and image datasets. These models achieve state-of-the-art results on downstream tasks, such as image captioning, visual question answering and open vocabulary recognition. Despite such achievements, these models require a massive volume of data for training and end up with a tremendous number of parameters (billions in many cases), resulting in significant computational requirements. Moreover, the data used to train these models can become outdated, requiring re-training every time the world’s knowledge is updated. For example, a model trained just two years ago might yield outdated information about the current president of the United States.

In the fields of natural language processing (RETRO, REALM) and computer vision (KAT), researchers have attempted to address these challenges using retrieval-augmented models. Typically, these models use a backbone that is able to process a single modality at a time, e.g., only text or only images, to encode and retrieve information from a knowledge corpus. However, these retrieval-augmented models are unable to leverage all available modalities in a query and knowledge corpora, and may not find the information that is most helpful for generating the model’s output.

To address these issues, in “REVEAL: Retrieval-Augmented Visual-Language Pre-Training with Multi-Source Multimodal Knowledge Memory”, to appear at CVPR 2023, we introduce a visual-language model that learns to utilize a multi-source multi-modal “memory” to answer knowledge-intensive queries. REVEAL employs neural representation learning to encode and convert diverse knowledge sources into a memory structure consisting of key-value pairs. The keys serve as indices for the memory items, while the corresponding values store pertinent information about those items. During training, REVEAL learns the key embeddings, value tokens, and the ability to retrieve information from this memory to address knowledge-intensive queries. This approach allows the model parameters to focus on reasoning about the query, rather than being dedicated to memorization.

We augment a visual-language model with the ability to retrieve multiple knowledge entries from a diverse set of knowledge sources, which helps generation.

Memory construction from multimodal knowledge corpora

Our approach is similar to REALM in that we precompute key and value embeddings of knowledge items from different sources and index them in a unified knowledge memory, where each knowledge item is encoded into a key-value pair. Each key is a d-dimensional embedding vector, while each value is a sequence of token embeddings representing the knowledge item in more detail. In contrast to previous work, REVEAL leverages a diverse set of multimodal knowledge corpora, including the WikiData knowledge graph, Wikipedia passages and images, web image-text pairs and visual question answering data. Each knowledge item could be text, an image, a combination of both (e.g., pages in Wikipedia) or a relationship or attribute from a knowledge graph (e.g., Barack Obama is 6’ 2” tall). During training, we continuously re-compute the memory key and value embeddings as the model parameters get updated. We update the memory asynchronously at every thousand training steps.

Scaling memory using compression

A naïve solution for encoding a memory value is to keep the whole sequence of tokens for each knowledge item. Then, the model could fuse the input query and the top-k retrieved memory values by concatenating all their tokens together and feeding them into a transformer encoder-decoder pipeline. This approach has two issues: (1) storing hundreds of millions of knowledge items in memory is impractical if each memory value consists of hundreds of tokens and (2) the transformer encoder has a quadratic complexity with respect to the total number of tokens times k for self-attention. Therefore, we propose to use the Perceiver architecture to encode and compress knowledge items. The Perceiver model uses a transformer decoder to compress the full token sequence into an arbitrary length. This lets us retrieve top-k memory entries for k as large as a hundred.

The following figure illustrates the procedure of constructing the memory key-value pairs. Each knowledge item is processed through a multi-modal visual-language encoder, resulting in a sequence of image and text tokens. The key head then transforms these tokens into a compact embedding vector. The value head (perceiver) condenses these tokens into fewer ones, retaining the pertinent information about the knowledge item within them.

We encode the knowledge entries from different corpora into unified key and value embedding pairs, where the keys are used to index the memory and values contain information about the entries.

Large-scale pre-training on image-text pairs

To train the REVEAL model, we begin with the large-scale corpus, collected from the public Web with three billion image alt-text caption pairs, introduced in LiT. Since the dataset is noisy, we add a filter to remove data points with captions shorter than 50 characters, which yields roughly 1.3 billion image caption pairs. We then take these pairs, combined with the text generation objective used in SimVLM, to train REVEAL. Given an image-text example, we randomly sample a prefix containing the first few tokens of the text. We feed the text prefix and image to the model as input with the objective of generating the rest of the text as output. The training goal is to condition the prefix and autoregressively generate the remaining text sequence.

To train all components of the REVEAL model end-to-end, we need to warm start the model to a good state (setting initial values to model parameters). Otherwise, if we were to start with random weights (cold-start), the retriever would often return irrelevant memory items that would never generate useful training signals. To avoid this cold-start problem, we construct an initial retrieval dataset with pseudo–ground-truth knowledge to give the pre-training a reasonable head start.

We create a modified version of the WIT dataset for this purpose. Each image-caption pair in WIT also comes with a corresponding Wikipedia passage (words surrounding the text). We put together the surrounding passage with the query image and use it as the pseudo ground-truth knowledge that corresponds to the input query. The passage provides rich information about the image and caption, which is useful for initializing the model.

To prevent the model from relying on low-level image features for retrieval, we apply random data augmentation to the input query image. Given this modified dataset that contains pseudo-retrieval ground-truth, we train the query and memory key embeddings to warm start the model.

REVEAL workflow

The overall workflow of REVEAL consists of four primary steps. First, REVEAL encodes a multimodal input into a sequence of token embeddings along with a condensed query embedding. Then, the model translates each multi-source knowledge entry into unified pairs of key and value embeddings, with the key being utilized for memory indexing and the value encompassing the entire information about the entry. Next, REVEAL retrieves the top-k most related knowledge pieces from multiple knowledge sources, returns the pre-processed value embeddings stored in memory, and re-encodes the values. Finally, REVEAL fuses the top-k knowledge pieces through an attentive knowledge fusion layer by injecting the retrieval score (dot product between query and key embeddings) as a prior during attention calculation. This structure is instrumental in enabling the memory, encoder, retriever and the generator to be concurrently trained in an end-to-end fashion.

Overall workflow of REVEAL.

Results

We evaluate REVEAL on knowledge-based visual question answering tasks using OK-VQA and A-OKVQA datasets. We fine-tune our pre-trained model on the VQA tasks using the same generative objective where the model takes in an image-question pair as input and generates the text answer as output. We demonstrate that REVEAL achieves better results on the A-OKVQA dataset than earlier attempts that incorporate a fixed knowledge or the works that utilize large language models (e.g., GPT-3) as an implicit source of knowledge.

Visual question answering results on A-OKVQA. REVEAL achieves higher accuracy in comparison to previous works including ViLBERT, LXMERT, ClipCap, KRISP and GPV-2.

We also evaluate REVEAL on the image captioning benchmarks using MSCOCO and NoCaps dataset. We directly fine-tune REVEAL on the MSCOCO training split via the cross-entropy generative objective. We measure our performance on the MSCOCO test split and NoCaps evaluation set using the CIDEr metric, which is based on the idea that good captions should be similar to reference captions in terms of word choice, grammar, meaning, and content. Our results on MSCOCO caption and NoCaps datasets are shown below.

Image Captioning results on MSCOCO and NoCaps using the CIDEr metric. REVEAL achieves a higher score in comparison to Flamingo, VinVL, SimVLM and CoCa.

Below we show a couple of qualitative examples of how REVEAL retrieves relevant documents to answer visual questions.

REVEAL can use knowledge from different sources to correctly answer the question.

Conclusion

We present an end-to-end retrieval-augmented visual language (REVEAL) model, which contains a knowledge retriever that learns to utilize a diverse set of knowledge sources with different modalities. We train REVEAL on a massive image-text corpus with four diverse knowledge corpora, and achieve state-of-the-art results on knowledge-intensive visual question answering and image caption tasks. In the future we would like to explore the ability of this model for attribution, and apply it to a broader class of multimodal tasks.

Acknowledgements

This research was conducted by Ziniu Hu, Ahmet Iscen, Chen Sun, Zirui Wang, Kai-Wei Chang, Yizhou Sun, Cordelia Schmid, David A. Ross and Alireza Fathi.

Posted by Petros Maniatis and Daniel Tarlow, Research Scientists, Google

Software isn’t created in one dramatic step. It improves bit by bit, one little step at a time — editing, running unit tests, fixing build errors, addressing code reviews, editing some more, appeasing linters, and fixing more errors — until finally it becomes good enough to merge into a code repository. Software engineering isn’t an isolated process, but a dialogue among human developers, code reviewers, bug reporters, software architects and tools, such as compilers, unit tests, linters and static analyzers.

Today we describe DIDACT (​​Dynamic Integrated Developer ACTivity), which is a methodology for training large machine learning (ML) models for software development. The novelty of DIDACT is that it uses the process of software development as the source of training data for the model, rather than just the polished end state of that process, the finished code. By exposing the model to the contexts that developers see as they work, paired with the actions they take in response, the model learns about the dynamics of software development and is more aligned with how developers spend their time. We leverage instrumentation of Google’s software development to scale up the quantity and diversity of developer-activity data beyond previous works. Results are extremely promising along two dimensions: usefulness to professional software developers, and as a potential basis for imbuing ML models with general software development skills.

DIDACT is a multi-task model trained on development activities that include editing, debugging, repair, and code review.

We built and deployed internally three DIDACT tools, Comment Resolution (which we recently announced), Build Repair, and Tip Prediction, each integrated at different stages of the development workflow. All three of these tools received enthusiastic feedback from thousands of internal developers. We see this as the ultimate test of usefulness: do professional developers, who are often experts on the code base and who have carefully honed workflows, leverage the tools to improve their productivity?

Perhaps most excitingly, we demonstrate how DIDACT is a first step towards a general-purpose developer-assistance agent. We show that the trained model can be used in a variety of surprising ways, via prompting with prefixes of developer activities, and by chaining together multiple predictions to roll out longer activity trajectories. We believe DIDACT paves a promising path towards developing agents that can generally assist across the software development process.

A treasure trove of data about the software engineering process

Google’s software engineering toolchains store every operation related to code as a log of interactions among tools and developers, and have done so for decades. In principle, one could use this record to replay in detail the key episodes in the “software engineering video” of how Google’s codebase came to be, step-by-step — one code edit, compilation, comment, variable rename, etc., at a time.

Google code lives in a monorepo, a single repository of code for all tools and systems. A software developer typically experiments with code changes in a local copy-on-write workspace managed by a system called Clients in the Cloud (CitC). When the developer is ready to package a set of code changes together for a specific purpose (e.g., fixing a bug), they create a changelist (CL) in Critique, Google’s code-review system. As with other types of code-review systems, the developer engages in a dialog with a peer reviewer about functionality and style. The developer edits their CL to address reviewer comments as the dialog progresses. Eventually, the reviewer declares “LGTM!” (“looks good to me”), and the CL is merged into the code repository.

Of course, in addition to a dialog with the code reviewer, the developer also maintains a “dialog” of sorts with a plethora of other software engineering tools, such as the compiler, the testing framework, linters, static analyzers, fuzzers, etc.

An illustration of the intricate web of activities involved in developing software: small actions by the developer, interactions with a code reviewer, and invocations of tools such as compilers.

A multi-task model for software engineering

DIDACT utilizes interactions among engineers and tools to power ML models that assist Google developers, by suggesting or enhancing actions developers take — in context — while pursuing their software-engineering tasks. To do that, we have defined a number of tasks about individual developer activities: repairing a broken build, predicting a code-review comment, addressing a code-review comment, renaming a variable, editing a file, etc. We use a common formalism for each activity: it takes some State (a code file), some Intent (annotations specific to the activity, such as code-review comments or compiler errors), and produces an Action (the operation taken to address the task). This Action is like a mini programming language, and can be extended for newly added activities. It covers things like editing, adding comments, renaming variables, marking up code with errors, etc. We call this language DevScript.

The DIDACT model is prompted with a task, code snippets, and annotations related to that task, and produces development actions, e.g., edits or comments.

This state-intent-action formalism enables us to capture many different tasks in a general way. What’s more, DevScript is a concise way to express complex actions, without the need to output the whole state (the original code) as it would be after the action takes place; this makes the model more efficient and more interpretable. For example, a rename might touch a file in dozens of places, but a model can predict a single rename action.

An ML peer programmer

DIDACT does a good job on individual assistive tasks. For example, below we show DIDACT doing code clean-up after functionality is mostly done. It looks at the code along with some final comments by the code reviewer (marked with “human” in the animation), and predicts edits to address those comments (rendered as a diff).

Given an initial snippet of code and the comments that a code reviewer attached to that snippet, the Pre-Submit Cleanup task of DIDACT produces edits (insertions and deletions of text) that address those comments.

The multimodal nature of DIDACT also gives rise to some surprising capabilities, reminiscent of behaviors emerging with scale. One such capability is history augmentation, which can be enabled via prompting. Knowing what the developer did recently enables the model to make a better guess about what the developer should do next.

An illustration of history-augmented code completion in action.

A powerful such task exemplifying this capability is history-augmented code completion. In the figure below, the developer adds a new function parameter (1), and moves the cursor into the documentation (2). Conditioned on the history of developer edits and the cursor position, the model completes the line (3) by correctly predicting the docstring entry for the new parameter.

An illustration of edit prediction, over multiple chained iterations.

In an even more powerful history-augmented task, edit prediction, the model can choose where to edit next in a fashion that is historically consistent. If the developer deletes a function parameter (1), the model can use history to correctly predict an update to the docstring (2) that removes the deleted parameter (without the human developer manually placing the cursor there) and to update a statement in the function (3) in a syntactically (and — arguably — semantically) correct way. With history, the model can unambiguously decide how to continue the “editing video” correctly. Without history, the model wouldn’t know whether the missing function parameter is intentional (because the developer is in the process of a longer edit to remove it) or accidental (in which case the model should re-add it to fix the problem).

The model can go even further. For example, we started with a blank file and asked the model to successively predict what edits would come next until it had written a full code file. The astonishing part is that the model developed code in a step-by-step way that would seem natural to a developer: It started by first creating a fully working skeleton with imports, flags, and a basic main function. It then incrementally added new functionality, like reading from a file and writing results, and added functionality to filter out some lines based on a user-provided regular expression, which required changes across the file, like adding new flags.

Conclusion

DIDACT turns Google’s software development process into training demonstrations for ML developer assistants, and uses those demonstrations to train models that construct code in a step-by-step fashion, interactively with tools and code reviewers. These innovations are already powering tools enjoyed by Google developers every day. The DIDACT approach complements the great strides taken by large language models at Google and elsewhere, towards technologies that ease toil, improve productivity, and enhance the quality of work of software engineers.

Acknowledgements

This work is the result of a multi-year collaboration among Google Research, Google Core Systems and Experiences, and DeepMind. We would like to acknowledge our colleagues Jacob Austin, Pascal Lamblin, Pierre-Antoine Manzagol, and Daniel Zheng, who join us as the key drivers of this project. This work could not have happened without the significant and sustained contributions of our partners at Alphabet (Peter Choy, Henryk Michalewski, Subhodeep Moitra, Malgorzata Salawa, Vaibhav Tulsyan, and Manushree Vijayvergiya), as well as the many people who collected data, identified tasks, built products, strategized, evangelized, and helped us execute on the many facets of this agenda (Ankur Agarwal, Paige Bailey, Marc Brockschmidt, Rodrigo Damazio Bovendorp, Satish Chandra, Savinee Dancs, Matt Frazier, Alexander Frömmgen, Nimesh Ghelani, Chris Gorgolewski, Chenjie Gu, Vincent Hellendoorn, Franjo Ivančić, Marko Ivanković, Emily Johnston, Luka Kalinovcic, Lera Kharatyan, Jessica Ko, Markus Kusano, Kathy Nix, Sara Qu, Marc Rasi, Marcus Revaj, Ballie Sandhu, Michael Sloan, Tom Small, Gabriela Surita, Maxim Tabachnyk, David Tattersall, Sara Toth, Kevin Villela, Sara Wiltberger, and Donald Duo Zhao) and our extremely supportive leadership (Martín Abadi, Joelle Barral, Jeff Dean, Madhura Dudhgaonkar, Douglas Eck, Zoubin Ghahramani, Hugo Larochelle, Chandu Thekkath, and Niranjan Tulpule). Thank you!

Posted by Lucas Dixon and Michael Terry, co-leads, PAIR, Google Research

PAIR (People + AI Research) first launched in 2017 with the belief that “AI can go much further — and be more useful to all of us — if we build systems with people in mind at the start of the process.” We continue to focus on making AI more understandable, interpretable, fun, and usable by more people around the world. It’s a mission that is particularly timely given the emergence of generative AI and chatbots.

Today, PAIR is part of the Responsible AI and Human-Centered Technology team within Google Research, and our work spans this larger research space: We advance foundational research on human-AI interaction (HAI) and machine learning (ML); we publish educational materials, including the PAIR Guidebook and Explorables (such as the recent Explorable looking at how and why models sometimes make incorrect predictions confidently); and we develop software tools like the Learning Interpretability Tool to help people understand and debug ML behaviors. Our inspiration this year is “changing the way people think about what THEY can do with AI.” This vision is inspired by the rapid emergence of generative AI technologies, such as large language models (LLMs) that power chatbots like Bard, and new generative media models like Google’s Imagen, Parti, and MusicLM. In this blog post, we review recent PAIR work that is changing the way we engage with AI.

Generative AI research

Generative AI is creating a lot of excitement, and PAIR is involved in a range of related research, from using language models to create generative agents to studying how artists adopted generative image models like Imagen and Parti. These latter “text-to-image” models let a person input a text-based description of an image for the model to generate (e.g., “a gingerbread house in a forest in a cartoony style”). In a forthcoming paper titled “The Prompt Artists” (to appear in Creativity and Cognition 2023), we found that users of generative image models strive not only to create beautiful images, but also to create unique, innovative styles. To help achieve these styles, some would even seek unique vocabulary to help develop their visual style. For example, they may visit architectural blogs to learn what domain-specific vocabulary they can adopt to help produce distinctive images of buildings.

We are also researching solutions to challenges faced by prompt creators who, with generative AI, are essentially programming without using a programming language. As an example, we developed new methods for extracting semantically meaningful structure from natural language prompts. We have applied these structures to prompt editors to provide features similar to those found in other programming environments, such as semantic highlighting, autosuggest, and structured data views.

The growth of generative LLMs has also opened up new techniques to solve important long-standing problems. Agile classifiers are one approach we’re taking to leverage the semantic and syntactic strengths of LLMs to solve classification problems related to safer online discourse, such as nimbly blocking newer types of toxic language as quickly as it may evolve online. The big advance here is the ability to develop high quality classifiers from very small datasets — as small as 80 examples. This suggests a positive future for online discourse and better moderation of it: instead of collecting millions of examples to attempt to create universal safety classifiers for all use cases over months or years, more agile classifiers might be created by individuals or small organizations and tailored for their specific use cases, and iterated on and adapted in the time-span of a day (e.g., to block a new kind of harassment being received or to correct unintended biases in models). As an example of their utility, these methods recently won a SemEval competition to identify and explain sexism.

We’ve also developed new state-of-the-art explainability methods to identify the role of training data on model behaviors and misbehaviours. By combining training data attribution methods with agile classifiers, we also found that we can identify mislabelled training examples. This makes it possible to reduce the noise in training data, leading to significant improvements on model accuracy.

Collectively, these methods are critical to help the scientific community improve generative models. They provide techniques for fast and effective content moderation and dialogue safety methods that help support creators whose content is the basis for generative models’ amazing outcomes. In addition, they provide direct tools to help debug model misbehavior which leads to better generation.

Visualization and education

To lower barriers in understanding ML-related work, we regularly design and publish highly visual, interactive online essays, called AI Explorables, that provide accessible, hands-on ways to learn about key ideas in ML. For example, we recently published new AI Explorables on the topics of model confidence and unintended biases. In our latest Explorable, “From Confidently Incorrect Models to Humble Ensembles,” we discuss the problem with model confidence: models can sometimes be very confident in their predictions… and yet completely incorrect. Why does this happen and what can be done about it? Our Explorable walks through these issues with interactive examples and shows how we can build models that have more appropriate confidence in their predictions by using a technique called ensembling, which works by averaging the outputs of multiple models. Another Explorable, “Searching for Unintended Biases with Saliency”, shows how spurious correlations can lead to unintended biases — and how techniques such as saliency maps can detect some biases in datasets, with the caveat that it can be difficult to see bias when it’s more subtle and sporadic in a training set.

PAIR designs and publishes AI Explorables, interactive essays on timely topics and new methods in ML research, such as “From Confidently Incorrect Models to Humble Ensembles,” which looks at how and why models offer incorrect predictions with high confidence, and how “ensembling” the outputs of many models can help avoid this.

Transparency and the Data Cards Playbook

Continuing to advance our goal of helping people to understand ML, we promote transparent documentation. In the past, PAIR and Google Cloud developed model cards. Most recently, we presented our work on Data Cards at ACM FAccT’22 and open-sourced the Data Cards Playbook, a joint effort with the Technology, AI, Society, and Culture team (TASC). The Data Cards Playbook is a toolkit of participatory activities and frameworks to help teams and organizations overcome obstacles when setting up a transparency effort. It was created using an iterative, multidisciplinary approach rooted in the experiences of over 20 teams at Google, and comes with four modules: Ask, Inspect, Answer and Audit. These modules contain a variety of resources that can help you customize Data Cards to your organization’s needs:

• 18 Foundations: Scalable frameworks that anyone can use on any dataset type
• 19 Transparency Patterns: Evidence-based guidance to produce high-quality Data Cards at scale
• 33 Participatory Activities: Cross-functional workshops to navigate transparency challenges for teams
• Interactive Lab: Generate interactive Data Cards from markdown in the browser

The Data Cards Playbook is accessible as a learning pathway for startups, universities, and other research groups.

Software Tools

Our team thrives on creating tools, toolkits, libraries, and visualizations that expand access and improve understanding of ML models. One such resource is Know Your Data, which allows researchers to test a model’s performance for various scenarios through interactive qualitative exploration of datasets that they can use to find and fix unintended dataset biases.

Recently, PAIR released a new version of the Learning Interpretability Tool (LIT) for model debugging and understanding. LIT v0.5 provides support for image and tabular data, new interpreters for tabular feature attribution, a “Dive” visualization for faceted data exploration, and performance improvements that allow LIT to scale to 100k dataset entries. You can find the release notes and code on GitHub.

PAIR’s Learning Interpretability Tool (LIT), an open-source platform for visualization and understanding of ML models.

PAIR has also contributed to MakerSuite, a tool for rapid prototyping with LLMs using prompt programming. MakerSuite builds on our earlier research on PromptMaker, which won an honorable mention at CHI 2022. MakerSuite lowers the barrier to prototyping ML applications by broadening the types of people who can author these prototypes and by shortening the time spent prototyping models from months to minutes. 

A screenshot of MakerSuite, a tool for rapidly prototyping new ML models using prompt-based programming, which grew out of PAIR’s prompt programming research.

Ongoing work

As the world of AI moves quickly ahead, PAIR is excited to continue to develop new tools, research, and educational materials to help change the way people think about what THEY can do with AI.

For example, we recently conducted an exploratory study with five designers (presented at CHI this year) that looks at how people with no ML programming experience or training can use prompt programming to quickly prototype functional user interface mock-ups. This prototyping speed can help inform designers on how to integrate ML models into products, and enables them to conduct user research sooner in the product design process.

Based on this study, PAIR’s researchers built PromptInfuser, a design tool plugin for authoring LLM-infused mock-ups. The plug-in introduces two novel LLM-interactions: input-output, which makes content interactive and dynamic, and frame-change, which directs users to different frames depending on their natural language input. The result is more tightly integrated UI and ML prototyping, all within a single interface.

Recent advances in AI represent a significant shift in how easy it is for researchers to customize and control models for their research objectives and goals.These capabilities are transforming the way we think about interacting with AI, and they create lots of new opportunities for the research community. PAIR is excited about how we can leverage these capabilities to make AI easier to use for more people.

Acknowledgements

Thanks to everyone in PAIR, to Reena Jana and to all of our collaborators.

Posted by Deborah Cohen, Staff Research Scientist, and Craig Boutilier, Principal Scientist, Google Research

As virtual assistants become ubiquitous, users increasingly interact with them to learn about new topics or obtain recommendations and expect them to deliver capabilities beyond narrow dialogues of one or two turns. Dynamic planning, namely the capability to look ahead and replan based on the flow of the conversation, is an essential ingredient for the making of engaging conversations with the deeper, open-ended interactions that users expect.

While large language models (LLMs) are now beating state-of-the-art approaches in many natural language processing benchmarks, they are typically trained to output the next best response, rather than planning ahead, which is required for multi-turn interactions. However, in the past few years, reinforcement learning (RL) has delivered incredible results addressing specific problems that involve dynamic planning, such as winning games and protein folding.

Today, we are sharing our recent advances in dynamic planning for human-to-assistant conversations, in which we enable an assistant to plan a multi-turn conversation towards a goal and adapt that plan in real-time by adopting an RL-based approach. Here we look at how to improve long interactions by applying RL to compose answers based on information extracted from reputable sources, rather than relying on content generated by a language model. We expect that future versions of this work could combine LLMs and RL in multi-turn dialogues. The deployment of RL “in the wild” in a large-scale dialogue system proved a formidable challenge due to the modeling complexity, tremendously large state and action spaces, and significant subtlety in designing reward functions.

What is dynamic planning?

Many types of conversations, from gathering information to offering recommendations, require a flexible approach and the ability to modify the original plan for the conversation based on its flow. This ability to shift gears in the middle of a conversation is known as dynamic planning, as opposed to static planning, which refers to a more fixed approach. In the conversation below, for example, the goal is to engage the user by sharing interesting facts about cool animals. To begin, the assistant steers the conversation to sharks via a sound quiz. Given the user’s lack of interest in sharks, the assistant then develops an updated plan and pivots the conversation to sea lions, lions, and then cheetahs.

The assistant dynamically modifies its original plan to talk about sharks and shares facts about other animals.

Dynamic composition

To cope with the challenge of conversational exploration, we separate the generation of assistant responses into two parts: 1) content generation, which extracts relevant information from reputable sources, and 2) flexible composition of such content into assistant responses. We refer to this two-part approach as dynamic composition. Unlike LLM methods, this approach gives the assistant the ability to fully control the source, correctness, and quality of the content that it may offer. At the same time, it can achieve flexibility via a learned dialogue manager that selects and combines the most appropriate content.

In an earlier paper, “Dynamic Composition for Conversational Domain Exploration”, we describe a novel approach which consists of: (1) a collection of content providers, which offer candidates from different sources, such as news snippets, knowledge graph facts, and questions; (2) a dialogue manager; and (3) a sentence fusion module. Each assistant response is incrementally constructed by the dialogue manager, which selects candidates proposed by the content providers. The selected sequence of utterances is then fused into a cohesive response.

Dynamic planning using RL

At the core of the assistant response composition loop is a dialogue manager trained using off-policy RL, namely an algorithm that evaluates and improves a policy that is different from the policy used by the agent (in our case, the latter is based on a supervised model). Applying RL to dialogue management presents several challenges, including a large state space (as the state represents the conversation state, which needs to account for the whole conversation history) and an effectively unbounded action space (that may include all existing words or sentences in natural language).

We address these challenges using a novel RL construction. First, we leverage powerful supervised models — specifically, recurrent neural networks (RNNs) and transformers — to provide a succinct and effective dialogue state representation. These state encoders are fed with the dialogue history, composed of a sequence of user and assistant turns, and output a representation of the dialogue state in the form of a latent vector.

Second, we use the fact that a relatively small set of reasonable candidate utterances or actions can be generated by content providers at each conversation turn, and limit the action space to these. Whereas the action space is typically fixed in RL settings, because all states share the same action space, ours is a non-standard space in which the candidate actions may differ with each state, since content providers generate different actions depending on the dialogue context. This puts us in the realm of stochastic action sets, a framework that formalizes cases where the set of actions available in each state is governed by an exogenous stochastic process, which we address using Stochastic Action Q-Learning, a variant of the Q-learning approach. Q-learning is a popular off-policy RL algorithm, which does not require a model of the environment to evaluate and improve the policy. We trained our model on a corpus of crowd-compute–rated conversations obtained using a supervised dialogue manager.

Given the current dialogue history and a new user query, content providers generate candidates from which the assistant selects one. This process runs in a loop, and at the end the selected utterances are fused into a cohesive response.

Reinforcement learning model evaluation

We compared our RL dialogue manager with a launched supervised transformer model in an experiment using Google Assistant, which conversed with users about animals. A conversation starts when a user triggers the experience by asking an animal-related query (e.g., “How does a lion sound?”). The experiment was conducted using an A/B testing protocol, in which a small percentage of Assistant users were randomly sampled to interact with our RL-based assistant while other users interacted with the standard assistant.

We found that the RL dialogue manager conducts longer, more engaging conversations. It increases conversation length by 30% while improving user engagement metrics. We see an increase of 8% in cooperative responses to the assistant’s questions — e.g., “Tell me about lions,” in response to “Which animal do you want to hear about next?” Although there is also a large increase in nominally “non-cooperative” responses (e.g., “No,” as a reply to a question proposing additional content, such as “Do you want to hear more?”), this is expected as the RL agent takes more risks by asking pivoting questions. While a user may not be interested in the conversational direction proposed by the assistant (e.g., pivoting to another animal), the user will often continue to engage in a dialogue about animals.

From the non-cooperative user response in the 3rd turn (“No.”) and the query “Make a dog sound,” in the 5th turn, the assistant recognizes that the user is mostly interested in animal sounds and modifies its plan, providing sounds and sound quizzes.

In addition, some user queries contain explicit positive (e.g., “Thank you, Google,” or “I’m happy.”) or negative (e.g., “Shut up,” or “Stop.”) feedback. While an order of magnitude fewer than other queries, they offer a direct measure of user (dis)satisfaction. The RL model increases explicit positive feedback by 32% and reduces negative feedback by 18%.

Learned dynamic planning characteristics and strategies

We observe several characteristics of the (unseen) RL plan to improve user engagement while conducting longer conversations. First, the RL-based assistant ends 20% more turns in questions, prompting the user to choose additional content. It also better harnesses content diversity, including facts, sounds, quizzes, yes/no questions, open questions, etc. On average, the RL assistant uses 26% more distinct content providers per conversation than the supervised model.

Two observed RL planning strategies are related to the existence of sub-dialogues with different characteristics. Sub-dialogues about animal sounds are poorer in content and exhibit entity pivoting at every turn (i.e., after playing the sound of a given animal, we can either suggest the sound of a different animal or quiz the user about other animal sounds). In contrast, sub-dialogues involving animal facts typically contain richer content and have greater conversation depth. We observe that RL favors the richer experience of the latter, selecting 31% more fact-related content. Lastly, when restricting analysis to fact-related dialogues, the RL assistant exhibits 60% more focus-pivoting turns, that is, conversational turns that change the focus of the dialogue.

Below, we show two example conversations, one conducted by the supervised model (left) and the second by the RL model (right), in which the first three user turns are identical. With a supervised dialogue manager, after the user declined to hear about “today’s animal”, the assistant pivots back to animal sounds to maximize the immediate user satisfaction. While the conversation conducted by the RL model begins identically, it exhibits a different planning strategy to optimize the overall user engagement, introducing more diverse content, such as fun facts.

In the left conversation, conducted by the supervised model, the assistant maximizes the immediate user satisfaction. The right conversation, conducted by the RL model, shows different planning strategies to optimize the overall user engagement.

Future research and challenges

In the past few years, LLMs trained for language understanding and generation have demonstrated impressive results across multiple tasks, including dialogue. We are now exploring the use of an RL framework to empower LLMs with the capability of dynamic planning so that they can dynamically plan ahead and delight users with a more engaging experience.

Acknowledgements

The work described is co-authored by: Moonkyung Ryu, Yinlam Chow, Orgad Keller, Ido Greenberg, Avinatan Hassidim, Michael Fink, Yossi Matias, Idan Szpektor and Gal Elidan. We would like to thank: Roee Aharoni, Moran Ambar, John Anderson, Ido Cohn, Mohammad Ghavamzadeh, Lotem Golany, Ziv Hodak, Adva Levin, Fernando Pereira, Shimi Salant, Shachar Shimoni, Ronit Slyper, Ariel Stolovich, Hagai Taitelbaum, Noam Velan, Avital Zipori and the CrowdCompute team led by Ashwin Kakarla. We thank Sophie Allweis for her feedback on this blogpost and Tom Small for the visualization.

Posted by Jerry Wei, Student Researcher, and Denny Zhou, Principal Scientist, Google Research

There have recently been tremendous advances in language models, partly because they can perform tasks with strong performance via in-context learning (ICL), a process whereby models are prompted with a few examples of input-label pairs before performing the task on an unseen evaluation example. In general, models’ success at in-context learning is enabled by:

• Their use of semantic prior knowledge from pre-training to predict labels while following the format of in-context examples (e.g., seeing examples of movie reviews with “positive sentiment” and “negative sentiment” as labels and performing sentiment analysis using prior knowledge).
• Learning the input-label mappings in context from the presented examples (e.g., finding a pattern that positive reviews should be mapped to one label, and negative reviews should be mapped to a different label).

In “Larger language models do in-context learning differently”, we aim to learn about how these two factors (semantic priors and input-label mappings) interact with each other in ICL settings, especially with respect to the scale of the language model that’s used. We investigate two settings to study these two factors — ICL with flipped labels (flipped-label ICL) and ICL with semantically-unrelated labels (SUL-ICL). In flipped-label ICL, labels of in-context examples are flipped so that semantic priors and input-label mappings disagree with each other. In SUL-ICL, labels of in-context examples are replaced with words that are semantically unrelated to the task presented in-context. We found that overriding prior knowledge is an emergent ability of model scale, as is the ability to learn in-context with semantically-unrelated labels. We also found that instruction tuning strengthens the use of prior knowledge more than it increases the capacity to learn input-label mappings.

An overview of flipped-label ICL and semantically-unrelated label ICL (SUL-ICL), compared with regular ICL, for a sentiment analysis task. Flipped-label ICL uses flipped labels, forcing the model to override semantic priors in order to follow the in-context examples. SUL-ICL uses labels that are not semantically related to the task, which means that models must learn input-label mappings in order to perform the task because they can no longer rely on the semantics of natural language labels.

Experiment design

For a diverse dataset mixture, we experiment on seven natural language processing (NLP) tasks that have been widely used: sentiment analysis, subjective/objective classification, question classification, duplicated-question recognition, entailment recognition, financial sentiment analysis, and hate speech detection. We test five language model families, PaLM, Flan-PaLM, GPT-3, InstructGPT, and Codex.

Flipped labels

In this experiment, labels of in-context examples are flipped, meaning that prior knowledge and input-label mappings disagree (e.g., sentences containing positive sentiment labeled as “negative sentiment”), thereby allowing us to study whether models can override their priors. In this setting, models that are able to override prior knowledge and learn input-label mappings in-context should experience a decrease in performance (since ground-truth evaluation labels are not flipped).

The ability to override semantic priors when presented with flipped in-context example labels emerges with model scale. Smaller models cannot flip predictions to follow flipped labels (performance only decreases slightly), while larger models can do so (performance decreases to well below 50%).

We found that when no labels are flipped, larger models have better performance than smaller models (as expected). But when we flip more and more labels, the performance of small models stays relatively flat, but large models experience large performance drops to well-below random guessing (e.g., 90% → 22.5% for code-davinci-002).

These results indicate that large models can override prior knowledge from pre-training when contradicting input-label mappings are presented in-context. Small models can’t do this, making this ability an emergent phenomena of model scale.

Semantically-unrelated labels

In this experiment, we replace labels with semantically-irrelevant ones (e.g., for sentiment analysis, we use “foo/bar” instead of “negative/positive”), which means that the model can only perform ICL by learning from input-label mappings. If a model mostly relies on prior knowledge for ICL, then its performance should decrease after this change since it will no longer be able to use semantic meanings of labels to make predictions. A model that can learn input–label mappings in-context, on the other hand, would be able to learn these semantically-unrelated mappings and should not experience a major drop in performance.

Small models rely more on semantic priors than large models do, as indicated by the greater decrease in performance for small models than for large models when using semantically-unrelated labels (i.e., targets) instead of natural language labels. For each plot, models are shown in order of increasing model size (e.g., for GPT-3 models, a is smaller than b, which is smaller than c).

Indeed, we see that using semantically-unrelated labels results in a greater performance drop for small models. This suggests that smaller models primarily rely on their semantic priors for ICL rather than learning from the presented input-label mappings. Large models, on the other hand, have the ability to learn input-label mappings in-context when the semantic nature of labels is removed.

We also find that including more in-context examples (i.e., exemplars) results in a greater performance improvement for large models than it does for small models, indicating that large models are better at learning from in-context examples than small models are.

In the SUL-ICL setup, larger models benefit more from additional examples than smaller models do.

Instruction tuning

Instruction tuning is a popular technique for improving model performance, which involves tuning models on various NLP tasks that are phrased as instructions (e.g., “Question: What is the sentiment of the following sentence, ‘This movie is great.’ Answer: Positive”). Since the process uses natural language labels, however, an open question is whether it improves the ability to learn input-label mappings or whether it strengthens the ability to recognize and apply semantic prior knowledge. Both of these would lead to an improvement in performance on standard ICL tasks, so it’s unclear which of these occur.

We study this question by running the same two setups as before, only this time we focus on comparing standard language models (specifically, PaLM) with their instruction-tuned variants (Flan-PaLM).

First, we find that Flan-PaLM is better than PaLM when we use semantically-unrelated labels. This effect is very prominent in small models, as Flan-PaLM-8B outperforms PaLM-8B by 9.6% and almost catches up to PaLM-62B. This trend suggests that instruction tuning strengthens the ability to learn input-label mappings, which isn’t particularly surprising.

Instruction-tuned language models are better at learning input–label mappings than pre-training–only language models are.

More interestingly, we saw that Flan-PaLM is actually worse than PaLM at following flipped labels, meaning that the instruction tuned models were unable to override their prior knowledge (Flan-PaLM models don’t reach below random guessing with 100% flipped labels, but PaLM models without instruction tuning can reach 31% accuracy in the same setting). These results indicate that instruction tuning must increase the extent to which models rely on semantic priors when they’re available.

Instruction-tuned models are worse than pre-training–only models at learning to override semantic priors when presented with flipped labels in-context.

Combined with the previous result, we conclude that although instruction tuning improves the ability to learn input-label mappings, it strengthens the usage of semantic prior knowledge more.

Conclusion

We examined the extent to which language models learn in-context by utilizing prior knowledge learned during pre-training versus input-label mappings presented in-context.

We first showed that large language models can learn to override prior knowledge when presented with enough flipped labels, and that this ability emerges with model scale. We then found that successfully doing ICL using semantically-unrelated labels is another emergent ability of model scale. Finally, we analyzed instruction-tuned language models and saw that instruction tuning improves the capacity to learn input-label mappings but also strengthens the use of semantic prior knowledge even more.

Future work

These results underscore how the ICL behavior of language models can change depending on their scale, and that larger language models have an emergent ability to map inputs to many types of labels, a form of reasoning in which input-label mappings can potentially be learned for arbitrary symbols. Future research could help provide insights on why these phenomena occur with respect to model scale.

Acknowledgements

This work was conducted by Jerry Wei, Jason Wei, Yi Tay, Dustin Tran, Albert Webson, Yifeng Lu, Xinyun Chen, Hanxiao Liu, Da Huang, Denny Zhou, and Tengyu Ma. We would like to thank Sewon Min and our fellow collaborators at Google Research for their advice and helpful discussions.