Category: Offsites

Posted by Mingxing Tan and Zihang Dai, Research Scientists, Google Research

As neural network models and training data size grow, training efficiency is becoming an important focus for deep learning. For example, GPT-3 demonstrates remarkable capability in few-shot learning, but it requires weeks of training with thousands of GPUs, making it difficult to retrain or improve. What if, instead, one could design neural networks that were smaller and faster, yet still more accurate?

In this post, we introduce two families of models for image recognition that leverage neural architecture search, and a principled design methodology based on model capacity and generalization. The first is EfficientNetV2 (accepted at ICML 2021), which consists of convolutional neural networks that aim for fast training speed for relatively small-scale datasets, such as ImageNet1k (with 1.28 million images). The second family is CoAtNet, which are hybrid models that combine convolution and self-attention, with the goal of achieving higher accuracy on large-scale datasets, such as ImageNet21 (with 13 million images) and JFT (with billions of images). Compared to previous results, our models are 4-10x faster while achieving new state-of-the-art 90.88% top-1 accuracy on the well-established ImageNet dataset. We are also releasing the source code and pretrained models on the Google AutoML github.

EfficientNetV2: Smaller Models and Faster Training
EfficientNetV2 is based upon the previous EfficientNet architecture. To improve upon the original, we systematically studied the training speed bottlenecks on modern TPUs/GPUs and found: (1) training with very large image sizes results in higher memory usage and thus is often slower on TPUs/GPUs; (2) the widely used depthwise convolutions are inefficient on TPUs/GPUs, because they exhibit low hardware utilization; and (3) the commonly used uniform compound scaling approach, which scales up every stage of convolutional networks equally, is sub-optimal. To address these issues, we propose both a training-aware neural architecture search (NAS), in which the training speed is included in the optimization goal, and a scaling method that scales different stages in a non-uniform manner.

The training-aware NAS is based on the previous platform-aware NAS, but unlike the original approach, which mostly focuses on inference speed, here we jointly optimize model accuracy, model size, and training speed. We also extend the original search space to include more accelerator-friendly operations, such as FusedMBConv, and simplify the search space by removing unnecessary operations, such as average pooling and max pooling, which are never selected by NAS. The resulting EfficientNetV2 networks achieve improved accuracy over all previous models, while being much faster and up to 6.8x smaller.

To further speed up the training process, we also propose an enhanced method of progressive learning, which gradually changes image size and regularization magnitude during training. Progressive training has been used in image classification, GANs, and language models. This approach focuses on image classification, but unlike previous approaches that often trade accuracy for improved training speed, can slightly improve the accuracy while also significantly reducing training time. The key idea in our improved approach is to adaptively change regularization strength, such as dropout ratio or data augmentation magnitude, according to the image size. For the same network, small image size leads to lower network capacity and thus requires weak regularization; vice versa, a large image size requires stronger regularization to combat overfitting.

Progressive learning for EfficientNetV2. Here we mainly focus on three types of regularizations: data augmentation, mixup, and dropout.

We evaluate the EfficientNetV2 models on ImageNet and a few transfer learning datasets, such as CIFAR-10/100, Flowers, and Cars. On ImageNet, EfficientNetV2 significantly outperforms previous models with about 5–11x faster training speed and up to 6.8x smaller model size, without any drop in accuracy.

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

CoAtNet: Fast and Accurate Models for Large-Scale Image Recognition
While EfficientNetV2 is still a typical convolutional neural network, recent studies on Vision Transformer (ViT) have shown that attention-based transformer models could perform better than convolutional neural networks on large-scale datasets like JFT-300M. Inspired by this observation, we further expand our study beyond convolutional neural networks with the aim of finding faster and more accurate vision models.

In “CoAtNet: Marrying Convolution and Attention for All Data Sizes”, we systematically study how to combine convolution and self-attention to develop fast and accurate neural networks for large-scale image recognition. Our work is based on an observation that convolution often has better generalization (i.e., the performance gap between training and evaluation) due to its inductive bias, while self-attention tends to have greater capacity (i.e., the ability to fit large-scale training data) thanks to its global receptive field. By combining convolution and self-attention, our hybrid models can achieve both better generalization and greater capacity.

Comparison between convolution, self-attention, and hybrid models. Convolutional models converge faster, ViTs have better capacity, while the hybrid models achieve both faster convergence and better accuracy.

We observe two key insights from our study: (1) depthwise convolution and self-attention can be naturally unified via simple relative attention, and (2) vertically stacking convolution layers and attention layers in a way that considers their capacity and computation required in each stage (resolution) is surprisingly effective in improving generalization, capacity and efficiency. Based on these insights, we have developed a family of hybrid models with both convolution and attention, named CoAtNets (pronounced “coat” nets). The following figure shows the overall CoAtNet network architecture:

Overall CoAtNet architecture. Given an input image with size HxW, we first apply convolutions in the first stem stage (S0) and reduce the size to H/2 x W/2. The size continues to reduce with each stage. Ln refers to the number of layers. Then, the early two stages (S1 and S2) mainly adopt MBConv building blocks consisting of depthwise convolution. The later two stages (S3 and S4) mainly adopt Transformer blocks with relative self-attention. Unlike the previous Transformer blocks in ViT, here we use pooling between stages, similar to Funnel Transformer. Finally, we apply a classification head to generate class prediction.

CoAtNet models consistently outperform ViT models and its variants across a number of datasets, such as ImageNet1K, ImageNet21K, and JFT. When compared to convolutional networks, CoAtNet exhibits comparable performance on a small-scale dataset (ImageNet1K) and achieves substantial gains as the data size increases (e.g. on ImageNet21K and JFT).

Comparison between CoAtNet and previous models after pre-training on the medium sized ImageNet21K dataset. Under the same model size, CoAtNet consistently outperforms both ViT and convolutional models. Noticeably, with only ImageNet21K, CoAtNet is able to match the performance of ViT-H pre-trained on JFT.

We also evaluated CoAtNets on the large-scale JFT dataset. To reach a similar accuracy target, CoAtNet trains about 4x faster than previous ViT models and more importantly, achieves a new state-of-the-art top-1 accuracy on ImageNet of 90.88%.

Comparison between CoAtNets and previous ViTs. ImageNet top-1 accuracy after pre-training on JFT dataset under different training budget. The four best models are trained on JFT-3B with about 3 billion images.

Conclusion and Future Work
In this post, we introduce two families of neural networks, named EfficientNetV2 and CoAtNet, which achieve state-of-the-art performance on image recognition. All EfficientNetV2 models are open sourced and the pretrained models are also available on the TFhub. CoAtNet models will also be open-sourced soon. We hope these new neural networks can benefit the research community and the industry. In the future we plan to further optimize these models and apply them to new tasks, such as zero-shot learning and self-supervised learning, which often require fast models with high capacity.

Acknowledgements
Special thanks to our co-authors Hanxiao Liu and Quoc Le. We also thank the Google Research, Brain Team and the open source contributors.

Posted by Vighnesh Birodkar, Research Software Engineer and Jonathan Huang, Research Scientist, Google Research

Instance segmentation is the task of grouping pixels in an image into instances of individual things, and identifying those things with a class label (countable objects such as people, animals, cars, etc., and assigning unique identifiers to each, e.g., car_1 and car_2). As a core computer vision task, it is critical to many downstream applications, such as self-driving cars, robotics, medical imaging, and photo editing. In recent years, deep learning has made significant strides in solving the instance segmentation problem with architectures like Mask R-CNN. However, these methods rely on collecting a large labeled instance segmentation dataset. But unlike bounding box labels, which can be collected in 7 seconds per instance with methods like Extreme clicking, collecting instance segmentation labels (called “masks”) can take up to 80 seconds per instance, an effort that is costly and creates a high barrier to entry for this research. And a related task, pantopic segmentation, requires even more labeled data.

The partially supervised instance segmentation setting, where only a small set of classes are labeled with instance segmentation masks and the remaining (majority of) classes are labeled only with bounding boxes, is an approach that has the potential to reduce the dependence on manually-created mask labels, thereby significantly lowering the barriers to developing an instance segmentation model. However this partially supervised approach also requires a stronger form of model generalization to handle novel classes not seen at training time—e.g., training with only animal masks and then tasking the model to produce accurate instance segmentations for buildings or plants. Further, naïve approaches, such as training a class-agnostic Mask R-CNN, while ignoring mask losses for any instances that don’t have mask labels, have not worked well. For example, on the typical “VOC/Non-VOC” benchmark, where one trains on masks for a subset of 20 classes in COCO (called “seen classes”) and is tested on the remaining 60 classes (called “unseen classes”), a typical Mask R-CNN with Resnet-50 backbone gets to only ~18% mask mAP (mean Average Precision, higher is better) on unseen classes, whereas when fully supervised it can achieve a much higher >34% mask mAP on the same set.

In “The surprising impact of mask-head architecture on novel class segmentation”, to be presented at ICCV 2021, we identify the main culprits for Mask R-CNN’s poor performance on novel classes and propose two easy-to-implement fixes (one training protocol fix, one mask-head architecture fix) that work in tandem to close the gap to fully supervised performance. We show that our approach applies generally to crop-then-segment models, i.e., a Mask R-CNN or Mask R-CNN-like architecture that computes a feature representation of the entire image and then subsequently passes per-instance crops to a second-stage mask prediction network—also called a mask-head network. Putting our findings together, we propose a Mask R-CNN–based model that improves over the current state-of-the-art by a significant 4.7% mask mAP without requiring more complex auxiliary loss functions, offline trained priors, or weight transfer functions proposed by previous work. We have also open sourced the code bases for two versions of the model, called Deep-MAC and Deep-MARC, and published a colab to interactively produce masks like the video demo below.

A demo of our model, DeepMAC, which learns to predict accurate masks, given user specified boxes, even on novel classes that were not seen at training time. Try it yourself in the colab. Image credits: Chris Briggs, Wikipedia and Europeana.

Impact of Cropping Methodology in Partially Supervised Settings
An important step of crop-then-segment models is cropping—Mask R-CNN is trained by cropping a feature map as well as the ground truth mask to a bounding box corresponding to each instance. These cropped features are passed to another neural network (called a mask-head network) that computes a final mask prediction, which is then compared against the ground truth crop in the mask loss function. There are two choices for cropping: (1) cropping directly to the ground truth bounding box of an instance, or (2) cropping to bounding boxes predicted by the model (called, proposals). At test time, cropping is always performed with proposals as ground truth boxes are not assumed to be available.

Cropping to ground truth boxes vs. cropping to proposals predicted by a model during training. Standard Mask R-CNN implementations use both types of crops, but we show that cropping exclusively to ground truth boxes yields significantly stronger performance on novel categories.

We consider a general family of Mask R-CNN–like architectures with one small, but critical difference from typical Mask R-CNN training setups: we crop using ground truth boxes (instead of proposal boxes) at training time.

Typical Mask R-CNN implementations pass both types of crops to the mask head. However, this choice has traditionally been considered an unimportant implementation detail, because it does not affect performance significantly in the fully supervised setting. In contrast, for partially supervised settings, we find that cropping methodology plays a significant role—while cropping exclusively to ground truth boxes during training doesn’t change the results significantly in the fully supervised setting, it has a surprising and dramatic positive impact in the partially supervised setting, performing significantly better on unseen classes.

Performance of Mask R-CNN on unseen classes when trained with either proposals and ground truth (the default) or with only ground truth boxes. Training mask heads with only ground truth boxes yields a significant boost to performance on unseen classes, upwards of 9% mAP. We report performance with the ResNet-101-FPN backbone.

Unlocking the Full Generalization Potential of the Mask Head
Even more surprisingly, the above approach unlocks a novel phenomenon—with cropping-to-ground truth enabled during training, the mask head of Mask R-CNN takes on a disproportionate role in the ability of the model to generalize to unseen classes. As an example, in the following figure, we compare models that all have cropping-to-ground-truth enabled, but different out-of-the-box mask-head architectures on a parking meter, cell phone, and pizza (classes unseen during training).

Mask predictions for unseen classes with four different mask-head architectures (from left to right: ResNet-4, ResNet-12, ResNet-20, Hourglass-20, where the number refers to the number of layers of the neural network). Despite never having seen masks from the ‘parking meter’, ‘pizza’ or ‘mobile phone’ class, the rightmost mask-head architecture can segment these classes correctly. From left to right, we show better mask-head architectures predicting better masks. Moreover, this difference is only apparent when evaluating on unseen classes — if we evaluate on seen classes, all four architectures exhibit similar performance.

Particularly notable is that these differences between mask-head architectures are not as obvious in the fully supervised setting. Incidentally, this may explain why previous works in instance segmentation have almost exclusively used shallow (i.e., low number of layers) mask heads, as there has been no benefit to the added complexity. Below we compare the mask mAP of three different mask-head architectures on seen versus unseen classes. All three models do equally well on the set of seen classes, but the deep hourglass mask heads stand out when applied to unseen classes. We find hourglass mask heads to be the best among the architectures we tried and we use hourglass mask heads with 50 or more layers to get the best results.

Performance of ResNet-4, Hourglass-10 and Hourglass-52 mask-head architectures on seen and unseen classes. There is a significant difference in performance on unseen classes, even though the performance on seen classes barely changes.

Finally, we show that our findings are general, holding for a variety of backbones (e.g., ResNet, SpineNet, Hourglass) and detector architectures including anchor-based and anchor-free detectors and even when there is no detector at all.

Putting It Together
To achieve the best result, we combined the above findings: We trained a Mask R-CNN model with cropping-to-ground-truth enabled and a deep Hourglass-52 mask head with a SpineNet backbone on high resolution images (1280×1280). We call this model Deep-MARC (Deep Mask heads Above R–CNN). Without using any offline training or other hand-crafted priors, Deep-MARC exceeds previous state-of-the-art models by > 4.5% (absolute) mask mAP. Demonstrating the general nature of this approach, we also see strong results with a CenterNet-based (as opposed to Mask R-CNN-based) model (called Deep-MAC), which also exceeds the previous state of the art.

Comparison of Deep-MAC and Deep-MARC to other partially supervised instance segmentation approaches like MaskX R-CNN, ShapeMask and CPMask.

Conclusion
We develop instance segmentation models that are able to generalize to classes that were not part of the training set. We highlight the role of two key ingredients that can be applied to any crop-then-segment model (such as Mask R-CNN): (1) cropping-to-ground truth boxes during training, and (2) strong mask-head architectures. While neither of these ingredients have a large impact on the classes for which masks are available during training, employing both leads to significant improvement on novel classes for which masks are not available during training. Moreover, these ingredients are sufficient for achieving state-of-the-art-performance on the partially-supervised COCO benchmark. Finally, our findings are general and may also have implications for related tasks, such as panoptic segmentation and pose estimation.

Acknowledgements
We thank our co-authors Zhichao Lu, Siyang Li, and Vivek Rathod. We thank David Ross and our anonymous ICCV reviewers for their comments which played a big part in improving this research.

Posted by Shan Yang, Software Engineer and Angjoo Kanazawa, Research Scientist, Google Research

Dancing is a universal language found in nearly all cultures, and is an outlet many people use to express themselves on contemporary media platforms today. The ability to dance by composing movement patterns that align to music beats is a fundamental aspect of human behavior. However, dancing is a form of art that requires practice. In fact, professional training is often required to equip a dancer with a rich repertoire of dance motions needed to create expressive choreography. While this process is difficult for people, it is even more challenging for a machine learning (ML) model, because the task requires the ability to generate a continuous motion with high kinematic complexity, while capturing the non-linear relationship between the movements and the accompanying music.

In “AI Choreographer: Music-Conditioned 3D Dance Generation with AIST++”, presented at ICCV 2021, we propose a full-attention cross-modal Transformer (FACT) model can mimic and understand dance motions, and can even enhance a person’s ability to choreograph dance. Together with the model, we released a large-scale, multi-modal 3D dance motion dataset, AIST++, which contains 5.2 hours of 3D dance motion in 1408 sequences, covering 10 dance genres, each including multi-view videos with known camera poses. Through extensive user studies on AIST++, we find that the FACT model outperforms recent state-of-the-art methods, both qualitatively and quantitatively.

We present a novel full-attention cross-modal transformer (FACT) network that can generate realistic 3D dance motion (right) conditioned on music and a new 3D dance dataset, AIST++ (left).

We generate the proposed 3D motion dataset from the existing AIST Dance Database — a collection of videos of dance with musical accompaniment, but without any 3D information. AIST contains 10 dance genres: Old School (Break, Pop, Lock and Waack) and New School (Middle Hip-Hop, LA-style Hip-Hop, House, Krump, Street Jazz and Ballet Jazz). Although it contains multi-view videos of dancers, these cameras are not calibrated.

For our purposes, we recovered the camera calibration parameters and the 3D human motion in terms of parameters used by the widely used SMPL 3D model. The resulting database, AIST++, is a large-scale, 3D human dance motion dataset that contains a wide variety of 3D motion, paired with music. Each frame includes extensive annotations:

• 9 views of camera intrinsic and extrinsic parameters;
• 17 COCO-format human joint locations in both 2D and 3D;
• 24 SMPL pose parameters along with the global scaling and translation.

The motions are equally distributed among all 10 dance genres, covering a wide variety of music tempos in beat per minute (BPM). Each genre of dance contains 85% basic movements and 15% advanced movements (longer choreographies freely designed by the dancers).

The AIST++ dataset also contains multi-view synchronized image data, making it useful for other research directions, such as 2D/3D pose estimation. To our knowledge, AIST++ is the largest 3D human dance dataset with 1408 sequences, 30 subjects and 10 dance genres, and with both basic and advanced choreographies.

An example of a 3D dance sequence in the AIST++ dataset. Left: Three views of the dance video from the AIST database. Right: Reconstructed 3D motion visualized in 3D mesh (top) and skeletons (bottom).

Because AIST is an instructional database, it records multiple dancers following the same choreography for different music with varying BPM, a common practice in dance. This posits a unique challenge in cross-modal sequence-to-sequence generation as the model needs to learn the one-to-many mapping between audio and motion. We carefully construct non-overlapping train and test subsets on AIST++ to ensure neither choreography nor music is shared across the subsets.

Full Attention Cross-Modal Transformer (FACT) Model
Using this data, we train the FACT model to generate 3D dance from music. The model begins by encoding seed motion and audio inputs using separate motion and audio transformers. The embeddings are then concatenated and sent to a cross-modal transformer, which learns the correspondence between both modalities and generates N future motion sequences. These sequences are then used to train the model in a self-supervised manner. All three transformers are jointly learned end-to-end. At test time, we apply this model in an autoregressive framework, where the predicted motion serves as the input to the next generation step. As a result, the FACT model is capable of generating long range dance motion frame-by-frame.

The FACT network takes in a music piece (Y) and a 2-second sequence of seed motion (X), then generates long-range future motions that correlate with the input music.

FACT involves three key design choices that are critical for producing realistic 3D dance motion from music.

1. All of the transformers use a full-attention mask, which can be more expressive than typical causal models because internal tokens have access to all inputs.
2. We train the model to predict N futures beyond the current input, instead of just the next motion. This encourages the network to pay more attention to the temporal context, and helps prevent the model from motion freezing or diverging after a few generation steps.
3. We fuse the two embeddings (motion and audio) early and employ a deep 12-layer cross-modal transformer module, which is essential for training a model that actually pays attention to the input music.

Results
We evaluate the performance based on three metrics:

Motion Quality: We calculate the Frechet Inception Distance (FID) between the real dance motion sequences in the AIST++ test set and 40 model generated motion sequences, each with 1200 frames (20 secs). We denote the FID based on the geometric and kinetic features as FID_g and FID_k, respectively.

Generation Diversity: Similar to prior work, to evaluate the model’s ability to generate divers dance motions, we calculate the average Euclidean distance in the feature space across 40 generated motions on the AIST++ test set, again comparing geometric feature space (Dist_g) and in the kinetic feature space (Dist_k).

Four different dance choreographies (right) generated using different music, but the same two second seed motion (left). The genres of the conditioning music are: Break, Ballet Jazz, Krump and Middle Hip-hop. The seed motion comes from hip-hop dance.

Motion-Music Correlation: Because there is no well-designed metric to measure the correlation between input music (music beats) and generated 3D motion (kinematic beats), we propose a novel metric, called Beat Alignment Score (BeatAlign).

Kinetic velocity (blue curve) and kinematic beats (green dotted line) of the generated dance motion, as well as the music beats (orange dotted line). The kinematic beats are extracted by finding local minima from the kinetic velocity curve.

Quantitative Evaluation
We compare the performance of FACT on each of these metrics to that of other state-of-the-art methods.

Compared to three recent state-of-the-art methods (Li et al., Dancenet, and Dance Revolution), the FACT model generates motions that are more realistic, better correlated with input music, and more diversified when conditioned on different music. *Note that the Li et al. generated motions are discontinuous, making the average kinetic feature distance abnormally high.

We also perceptually evaluate the motion-music correlation with a user study in which each participant is asked to watch 10 videos showing one of our results and one random counterpart, and then select which dancer is more in sync with the music. The study consisted of 30 participants, ranging from professional dancers to people who rarely dance. Compared to each baseline, 81% prefered the FACT model output to that of Li et al., 71% prefered FACT to Dancenet, and 77% prefered it Dance Revolution. Interestingly, 75% of participants preferred the unpaired AIST++ dance motion to that generated by FACT, which is unsurprising since the original dance captures are highly expressive.

Qualitative Results
Compared with prior methods like DanceNet (left) and Li et. al. (middle), 3D dance generated using the FACT model (right) is more realistic and better correlated with input music.

More generated 3D dances using the FACT model.

Conclusion and Discussion
We present a model that can not only learn the audio-motion correspondence, but also can generate high quality 3D motion sequences conditioned on music. Because generating 3D movement from music is a nascent area of study, we hope our work will pave the way for future cross-modal audio to 3D motion generation. We are also releasing AIST++, the largest 3D human dance dataset to date. This proposed, multi-view, multi-genre, cross-modal 3D motion dataset can not only help research in the conditional 3D motion generation research but also human understanding research in general. We are releasing the code in our GitHub repository and the trained model here.

While our results show a promising direction in this problem of music conditioned 3D motion generation, there are more to be explored. First, our approach is kinematic-based and we do not reason about physical interactions between the dancer and the floor. Therefore the global translation can lead to artifacts, such as foot sliding and floating. Second, our model is currently deterministic. Exploring how to generate multiple realistic dances per music is an exciting direction.

Acknowledgements
We gratefully acknowledge the contribution of other co-authors, including Ruilong Li and David Ross. We thank Chen Sun, Austin Myers, Bryan Seybold and Abhijit Kundu for helpful discussions. We thank Emre Aksan and Jiaman Li for sharing their code. We also thank Kevin Murphy for the early attempts in this direction, as well as Peggy Chi and Pan Chen for the help on user study experiments.

Posted by Katrin Tomanek, Software Engineer and Bob MacDonald, Technical Program Manager, Google Research

Speech impairments affect millions of people, with underlying causes ranging from neurological or genetic conditions to physical impairment, brain damage or hearing loss. Similarly, the resulting speech patterns are diverse, including stuttering, dysarthria, apraxia, etc., and can have a detrimental impact on self-expression, participation in society and access to voice-enabled technologies. Automatic speech recognition (ASR) technologies have the potential to help individuals with such speech impairments by improving access to dictation and home automation and by enhancing communication. However, while the increased computational power of deep learning systems and the availability of large training datasets has improved the accuracy of ASR systems, their performance is still insufficient for many people with speech disorders, rendering the technology unusable for many of the speakers who could benefit the most.

In 2019, we introduced Project Euphonia and discussed how we could use personalized ASR models of disordered speech to achieve accuracies on par with non-personalized ASR on typical speech. Today we share the results of two studies, presented at Interspeech 2021, that aim to expand the availability of personalized ASR models to more users. In “Disordered Speech Data Collection: Lessons Learned at 1 Million Utterances from Project Euphonia”, we present a greatly expanded collection of disordered speech data, composed of over 1 million utterances. Then, in “Automatic Speech Recognition of Disordered Speech: Personalized models outperforming human listeners on short phrases”, we discuss our efforts to generate personalized ASR models based on this corpus. This approach leads to highly accurate models that can achieve up to 85% improvement to the word error rate in select domains compared to out-of-the-box speech models trained on typical speech.

Impaired Speech Data Collection
Since 2019, speakers with speech impairments of varying degrees of severity across a variety of conditions have provided voice samples to support Project Euphonia’s research mission. This effort has grown Euphonia’s corpus to over 1 million utterances, comprising over 1400 hours from 1330 speakers (as of August 2021).

Distribution of severity of speech disorder and condition across all speakers with more than 300 utterances recorded. For conditions, only those with > 5 speakers are shown (all others aggregated into “OTHER” for k-anonymity). ALS = amyotrophic lateral sclerosis; DS = Down syndrome; PD = Parkinson’s disease; CP = cerebral palsy; HI = hearing impaired; MD = muscular dystrophy; MS = multiple sclerosis

To simplify the data collection, participants used an at-home recording system on their personal hardware (laptop or phone, with and without headphones), instead of an idealized lab-based setting that would collect studio quality recordings.

To reduce transcription cost, while still maintaining high transcript conformity, we prioritized scripted speech. Participants read prompts shown on a browser-based recording tool. Phrase prompts covered use-cases like home automation (“Turn on the TV.”), caregiver conversations (“I am hungry.”) and informal conversations (“How are you doing? Did you have a nice day?”). Most participants received a list of 1500 phrases, which included 1100 unique phrases along with 100 phrases that were each repeated four more times.

Speech professionals conducted a comprehensive auditory-perceptual speech assessment while listening to a subset of utterances for every speaker providing the following speaker-level metadata: speech disorder type (e.g., stuttering, dysarthria, apraxia), rating of 24 features of abnormal speech (e.g., hypernasality, articulatory imprecision, dysprosody), as well as recording quality assessments of both technical (e.g., signal dropouts, segmentation problems) and acoustic (e.g., environmental noise, secondary speaker crosstalk) features.

Personalized ASR Models
This expanded impaired speech dataset is the foundation of our new approach to personalized ASR models for disordered speech. Each personalized model uses a standard end-to-end, RNN-Transducer (RNN-T) ASR model that is fine-tuned using data from the target speaker only.

Architecture of RNN-Transducer. In our case, the encoder network consists of 8 layers and the predictor network consists of 2 layers of uni-directional LSTM cells.

To accomplish this, we focus on adapting the encoder network, i.e. the part of the model dealing with the specific acoustics of a given speaker, as speech sound disorders were most common in our corpus. We found that only updating the bottom five (out of eight) encoder layers while freezing the top three encoder layers (as well as the joint layer and decoder layers) led to the best results and effectively avoided overfitting. To make these models more robust against background noise and other acoustic effects, we employ a configuration of SpecAugment specifically tuned to the prevailing characteristics of disordered speech. Further, we found that the choice of the pre-trained base model was critical. A base model trained on a large and diverse corpus of typical speech (multiple domains and acoustic conditions) proved to work best for our scenario.

Results
We trained personalized ASR models for ~430 speakers who recorded at least 300 utterances. 10% of utterances were held out as a test set (with no phrase overlap) on which we calculated the word error rate (WER) for the personalized model and the unadapted base model.

Overall, our personalization approach yields significant improvements across all severity levels and conditions. Even for severely impaired speech, the median WER for short phrases from the home automation domain dropped from around 89% to 13%. Substantial accuracy improvements were also seen across other domains such as conversational and caregiver.

WER of unadapted and personalized ASR models on home automation phrases.

To understand when personalization does not work well, we analyzed several subgroups:

• HighWER and LowWER: Speakers with high and low personalized model WERs based on the 1st and 5th quintiles of the WER distribution.
• SurpHighWER: Speakers with a surprisingly high WER (participants with typical speech or mild speech impairment of the HighWER group).

Different pathologies and speech disorder presentations are expected to impact ASR non-uniformly. The distribution of speech disorder types within the HighWER group indicates that dysarthria due to cerebral palsy was particularly difficult to model. Not surprisingly, median severity was also higher in this group.

To identify the speaker-specific and technical factors that impact ASR accuracy, we examined the differences (Cohen’s D) in the metadata between the participants that had poor (HighWER) and excellent (LowWER) ASR performance. As expected, overall speech severity was significantly lower in the LowWER group than in the HighWER group (p < 0.01). Intelligibility and severity were the most prominent atypical speech features in the HighWER group; however, other speech features also emerged, including abnormal prosody, articulation, and phonation. These speech features are known to degrade overall speech intelligibility.

The SurpHighWER group had fewer training utterances and lower SNR compared with the LowWER group (p < 0.01) resulting in large (negative) effect sizes, with all other factors having small effect sizes, except fastness. In contrast, the HighWER group exhibited medium to large differences across all factors.

Speech disorder and technical metadata effect sizes for the HighWER-vs-LowWER and SurpHighWER-vs-LowWER pairs. Positive effects indicated that the group values of the HighWER group were greater than LowWER groups.

We then compared personalized ASR models to human listeners. Three speech professionals independently transcribed 30 utterances per speaker. We found that WERs were, on average, lower for personalized ASR models compared to the WERs of human listeners, with gains increasing by severity.

Delta between the WERs of the personalized ASR models and the human listeners. Negative values indicate that personalized ASR performs better than human (expert) listeners.

Conclusions
With over 1 million utterances, Euphonia’s corpus is one of the largest and most diversely disordered speech corpora (in terms of disorder types and severities) and has enabled significant advances in ASR accuracy for these types of atypical speech. Our results demonstrate the efficacy of personalized ASR models for recognizing a wide range of speech impairments and severities, with potential for making ASR available to a wider population of users.

Acknowledgements
Key contributors to this project include Michael Brenner, Julie Cattiau, Richard Cave, Jordan Green, Rus Heywood, Pan-Pan Jiang, Anton Kast, Marilyn Ladewig, Bob MacDonald, Phil Nelson, Katie Seaver, Jimmy Tobin, and Katrin Tomanek. We gratefully acknowledge the support Project Euphonia received from members of many speech research teams across Google, including Françoise Beaufays, Fadi Biadsy, Dotan Emanuel, Khe Chai Sim, Pedro Moreno Mengibar, Arun Narayanan, Hasim Sak, Suzan Schwartz, Joel Shor, and many others. And most importantly, we wanted to say a huge thank you to the over 1300 participants who recorded speech samples and the many advocacy groups who helped us connect with these participants.

Posted by Kihyuk Sohn and Chun-Liang Li, Research Scientists, Google Cloud

Anomaly detection (sometimes called outlier detection or out-of-distribution detection) is one of the most common machine learning applications across many domains, from defect detection in manufacturing to fraudulent transaction detection in finance. It is most often used when it is easy to collect a large amount of known-normal examples but where anomalous data is rare and difficult to find. As such, one-class classification, such as one-class support vector machine (OC-SVM) or support vector data description (SVDD), is particularly relevant to anomaly detection because it assumes the training data are all normal examples, and aims to identify whether an example belongs to the same distribution as the training data. Unfortunately, these classical algorithms do not benefit from the representation learning that makes machine learning so powerful. On the other hand, substantial progress has been made in learning visual representations from unlabeled data via self-supervised learning, including rotation prediction and contrastive learning. As such, combining one-class classifiers with these recent successes in deep representation learning is an under-explored opportunity for the detection of anomalous data.

In “Learning and Evaluating Representations for Deep One-class Classification”, presented at ICLR 2021, we outline a 2-stage framework that makes use of recent progress on self-supervised representation learning and classic one-class algorithms. The algorithm is simple to train and results in state-of-the-art performance on various benchmarks, including CIFAR, f-MNIST, Cat vs Dog and CelebA. We then follow up on this in “CutPaste: Self-Supervised Learning for Anomaly Detection and Localization”, presented at CVPR 2021, in which we propose a new representation learning algorithm under the same framework for a realistic industrial defect detection problem. The framework achieves a new state-of-the-art on the MVTec benchmark.

A Two-Stage Framework for Deep One-Class Classification
While end-to-end learning has demonstrated success in many machine learning problems, including deep learning algorithm designs, such an approach for deep one-class classifiers often suffer from degeneration in which the model outputs the same results regardless of the input.

To combat this, we apply a two stage framework. In the first stage, the model learns deep representations with self-supervision. In the second stage, we adopt one-class classification algorithms, such as OC-SVM or kernel density estimator, using the learned representations from the first stage. This 2-stage algorithm is not only robust to degeneration, but also enables one to build more accurate one-class classifiers. Furthermore, the framework is not limited to specific representation learning and one-class classification algorithms — that is, one can easily plug-and-play different algorithms, which is useful if any advanced approaches are developed.

A deep neural network is trained to generate the representations of input images via self-supervision. We then train one-class classifiers on the learned representations.

Semantic Anomaly Detection
We test the efficacy of our 2-stage framework for anomaly detection by experimenting with two representative self-supervised representation learning algorithms, rotation prediction and contrastive learning.

Rotation prediction refers to a model’s ability to predict the rotated angles of an input image. Due to its promising performance in other computer vision applications, the end-to-end trained rotation prediction network has been widely adopted for one-class classification research. The existing approach typically reuses the built-in rotation prediction classifier for learning representations to conduct anomaly detection, which is suboptimal because those built-in classifiers are not trained for one-class classification.

In contrastive learning, a model learns to pull together representations from transformed versions of the same image, while pushing representations of different images away. During training, as images are drawn from the dataset, each is transformed twice with simple augmentations (e.g., random cropping or color changing). We minimize the distance of the representations from the same image to encourage consistency and maximize the distance between different images. However, usual contrastive learning converges to a solution where all the representations of normal examples are uniformly spread out on a sphere. This is problematic because most of the one-class algorithms determine the outliers by checking the proximity of a tested example to the normal training examples, but when all the normal examples are uniformly distributed in an entire space, outliers will always appear close to some normal examples.

To resolve this, we propose distribution augmentation (DA) for one-class contrastive learning. The idea is that instead of learning representations from the training data only, the model learns from the union of the training data plus augmented training examples, where the augmented examples are considered to be different from the original training data. We employ geometric transformations, such as rotation or horizontal flip, for distribution augmentation. With DA, the training data is no longer uniformly distributed in the representation space because some areas are occupied by the augmented data.

Left: Illustrated examples of perfect uniformity from the standard contrastive learning. Right: The reduced uniformity by the proposed distribution augmentation (DA), where the augmented data occupy the space to avoid the uniform distribution of the inlier examples (blue) throughout the whole sphere.

We evaluate the performance of one-class classification in terms of the area under receiver operating characteristic curve (AUC) on the commonly used datasets in computer vision, including CIFAR10 and CIFAR-100, Fashion MNIST, and Cat vs Dog. Images from one class are given as inliers and those from remaining classes are given as outliers. For example, we see how well cat images are detected as anomalies when dog images are inliers.

	CIFAR-10	CIFAR-100	f-MNIST	Cat v.s. Dog
Ruff et al. (2018)	64.8	–	–	–
Golan and El-Yaniv (2018)	86.0	78.7	93.5	88.8
Bergman and Hoshen (2020)	88.2	–	94.1	–
Hendrycks et al. (2019)	90.1	–	–	–
Huang et al. (2019)	86.6	78.8	93.9	–
2-stage framework: rotation prediction	91.3±0.3	84.1±0.6	95.8±0.3	86.4±0.6
2-stage framework: contrastive (DA)	92.5±0.6	86.5±0.7	94.8±0.3	89.6±0.5

Performance comparison of one-class classification methods. Values are the mean AUCs and their standard deviation over 5 runs. AUC ranges from 0 to 100, where 100 is perfect detection.

Given the suboptimal built-in rotation prediction classifiers typically used for rotation prediction approaches, it’s notable that simply replacing the built-in rotation classifier used in the first stage for learning representations with a one-class classifier at the second stage of the proposed framework significantly boosts the performance, from 86 to 91.3 AUC. More generally, the 2-stage framework achieves state-of-the-art performance on all of the above benchmarks.

With classic OC-SVM, which learns the area boundary of representations of normal examples, the 2-stage framework results in higher performance than existing works as measured by image-level AUC.

Texture Anomaly Detection for Industrial Defect Detection
In many real-world applications of anomaly detection, the anomaly is often defined by localized defects instead of entirely different semantics (i.e., being different in general). For example, the detection of texture anomalies is useful for detecting various kinds of industrial defects.

The examples of semantic anomaly detection and defect detection. In semantic anomaly detection, the inlier and outlier are different in general, (e.g., one is a dog, the other a cat). In defect detection, the semantics for inlier and outlier are the same (e.g., they are both tiles), but the outlier has a local anomaly.

While learning representations with rotation prediction and distribution-augmented contrastive learning have demonstrated state-of-the-art performance on semantic anomaly detection, those algorithms do not perform well on texture anomaly detection. Instead, we explored different representation learning algorithms that better fit the application.

In our second paper, we propose a new self-supervised learning algorithm for texture anomaly detection. The overall anomaly detection follows the 2-stage framework, but the first stage, in which the model learns deep image representations, is specifically trained to predict whether the image is augmented via a simple CutPaste data augmentation. The idea of CutPaste augmentation is simple — a given image is augmented by randomly cutting a local patch and pasting it back to a different location of the same image. Learning to distinguish normal examples from CutPaste-augmented examples encourages representations to be sensitive to local irregularity of an image.

The illustration of learning representations by predicting CutPaste augmentations. Given an example, the CutPaste augmentation crops a local patch, then pasties it to a randomly selected area of the same image. We then train a binary classifier to distinguish the original image and the CutPaste augmented image.

We use MVTec, a real-world defect detection dataset with 15 object categories, to evaluate the approach above.

DOCC (Ruff et al., 2020)	U-Student (Bergmann et al., 2020)	Rotation Prediction	Contrastive (DA)	CutPaste
87.9	92.5	86.3	86.5	95.2

Image-level anomaly detection performance (in AUC) on the MVTec benchmark.

Besides image-level anomaly detection, we use the CutPaste method to locate where the anomaly is, i.e., “patch-level” anomaly detection. We aggregate the patch anomaly scores via upsampling with Gaussian smoothing and visualize them in heatmaps that show where the anomaly is. Interestingly, this provides decently improved localization of anomalies. The below table shows the pixel-level AUC for localization evaluation.

Autoencoder (Bergmann et al., 2019)	FCDD (Ruff et al., 2020)	Rotation Prediction	Contrastive (DA)	CutPaste
86.0	92.0	93.0	90.4	96.0

Pixel-level anomaly localization performance (in AUC) comparison between different algorithms on the MVTec benchmark.

Conclusion
In this work we introduce a novel 2-stage deep one-class classification framework and emphasize the importance of decoupling building classifiers from learning representations so that the classifier can be consistent with the target task, one-class classification. Moreover, this approach permits applications of various self-supervised representation learning methods, attaining state-of-the-art performance on various applications of visual one-class classification from semantic anomaly to texture defect detection. We are extending our efforts to build more realistic anomaly detection methods under the scenario where training data is truly unlabeled.

Acknowledgements
We gratefully acknowledge the contribution from other co-authors, including Jinsung Yoon, Minho Jin and Tomas Pfister. We release the code in our GitHub repository.

Posted by Zaid Nabulsi, Software Engineer and Po-Hsuan Cameron Chen, Software Engineer, Google Health

The adoption of machine learning (ML) for medical imaging applications presents an exciting opportunity to improve the availability, latency, accuracy, and consistency of chest X-ray (CXR) image interpretation. Indeed, a plethora of algorithms have already been developed to detect specific conditions, such as lung cancer, tuberculosis and pneumothorax. By virtue of being trained to detect a specific disease, however, the utility of these algorithms may be limited in a general clinical setting, where a wide variety of abnormalities could surface. For example, a pneumothorax detector is not expected to highlight nodules suggestive of cancer, and a tuberculosis detector may not identify findings specific to pneumonia. Since an initial triaging step is to determine whether a CXR contains any concerning abnormalities, a general-purpose algorithm that identifies X-rays containing any sort of abnormality could significantly facilitate the workflow. However, developing a classifier to detect any abnormality is challenging due to the ​​wide variety of abnormal findings that present on CXRs.

In “Deep Learning for Distinguishing Normal versus Abnormal Chest Radiographs and Generalization to Two Unseen Diseases Tuberculosis and COVID-19”, published in Scientific Reports, we present a model that can distinguish between normal and abnormal CXRs across multiple de-identified datasets and settings. We find that the model performs well on general abnormalities, as well as unseen examples of tuberculosis and COVID-19. We are also releasing our set of radiologists’ labels¹ for the test set used in this study for the publicly available ChestX-ray14 dataset.

A Deep Learning System for Detecting Abnormal Chest X-rays
The deep learning system we used is based on the EfficientNet-B7 architecture, pre-trained on ImageNet. We trained the model using over 200,000 de-identified CXRs from the Apollo Hospitals in India. Each CXR was assigned a label of either “normal” or “abnormal” using a regular expression–based natural language processing approach on the associated radiology reports.

To evaluate how well the system generalizes to new patient populations, we compared its performance on two datasets consisting of a wide spectrum of abnormalities: the test split from the Apollo Hospitals dataset (DS-1), and the publicly available ChestX-ray14 (CXR-14). The labels for these two test sets were annotated for the purposes of this project by a group of US board-certified radiologists. The system achieved areas under the receiver operating characteristic curve (AUROC) of 0.87 on DS-1 and 0.94 on CXR-14 (higher is better).

Though the evaluations on DS-1 and CXR-14 contained a wide range of abnormalities, a possible use-case would be to utilize such an abnormality detector in novel or unforeseen settings with diseases that it had not encountered before. To evaluate the generalizability of the system to new patient populations and in the presence of diseases not seen in the training set, we used four de-identified datasets from three countries, including two publicly available tuberculosis datasets and two COVID-19 datasets from Northwestern Medicine. The system achieved AUCs of 0.95-0.97 in detecting tuberculosis, and 0.65-0.68 in detecting COVID-19. Because CXRs that are negative for these diseases could still contain other concerning abnormalities, we further evaluated the system for its ability to detect abnormalities more broadly (instead of disease positive vs. negative), finding AUCs of 0.91-0.93 for the tuberculosis dataset, and AUCs of 0.86 for the COVID-19 dataset.

The purpose of multiple evaluations (abnormality detection and disease detection) is the distinction between the two: a given disease can present with a certain abnormality or not; and a certain abnormality can arise from multiple diseases. Our study evaluates for both.

<!–

​​AUCs for Three Evaluation Setups
	1. General Abnormalities	2. Unseen disease: Tuberculosis	3. Unseen disease: COVID-19
Detect abnormalities	0.87-0.94	0.91-0.93	0.86
Detect respective disease	–	0.95-0.97	0.65-0.68

–>

The large drop in performance for COVID-19 is because many cases flagged by the system as “positive” for abnormalities were negative for COVID-19, but nevertheless contained abnormal CXR findings that needed attention. This further highlights the usefulness of abnormality detectors even if disease-specific models are available.

In addition, it’s important to note that there is a difference between generalization to unseen diseases (i.e., tuberculosis and COVID-19) versus generalization to unseen CXR findings (e.g., pleural effusion, consolidation/infiltrate). In this study, we demonstrated the generalizability of the system to unseen diseases but not necessarily unseen CXR findings.

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

Potential Benefits in the Clinic
To understand the potential utility of the deep learning model in improving clinical workflow, we simulated its use for case prioritization, where abnormal cases are “expedited” ahead of normal cases. In these simulations, the system reduced the turnaround time for abnormal cases by up to 28%. This reprioritization setup could be used to divert complex abnormal cases to cardiothoracic specialist radiologists, enable rapid triage of cases that may need urgent decisions, and provide the opportunity to batch negative CXRs for streamlined review.

Impact of a simulated deep learning model–based prioritization in comparison with random review order for (A) general abnormalities, (B) tuberculosis, and (C) COVID-19. The red bars indicate sequences of abnormal CXRs in red and normal CXRs in pink; a greater density of red towards the left indicates abnormal CXRs are reviewed sooner than normal ones. The histograms indicate the average improvement in turnaround time.

Additionally, we found that the system can be used as a pre-trained model to improve other ML algorithms for chest X-rays, especially when data is limited. For example, we used the normal/abnormal classifier in our recent study to detect pulmonary tuberculosis from chest X-rays. Abnormality and tuberculosis detectors can play a critical role in supporting early diagnosis in regions that lack access to resources like trained radiologists or molecular testing.

Sharing Improved Reference Standard Labels
Much work remains to be done to realize the potential of ML to aid chest X-ray interpretation around the world. In particular, obtaining high-quality labels on de-identified data can be a significant barrier to developing and evaluating ML algorithms in healthcare. To accelerate these efforts, we are expanding upon our previous label release by releasing the labels used in this study for the publicly available ChestX-ray14 dataset. We look forward to future machine learning projects by the community in this space.

AcknowledgementsKey contributors to this project at Google include Zaid Nabulsi, Andrew Sellergren‎, Shahar Jamshy, Charles Lau, Eddie Santos, Atilla P. Kiraly, Wenxing Ye, Jie Yang, Rory Pilgrim, Sahar Kazemzadeh, Jin Yu, Greg S. Corrado, Lily Peng, Krish Eswaran, Daniel Tse, Neeral Beladia, Yun Liu, Po-Hsuan Cameron Chen, Shravya Shetty. Significant contributions and input were also made by radiologist collaborators Sreenivasa Raju Kalidindi, Mozziyar Etemadi, Florencia Garcia Vicente, David Melnick. For the CXR-14 dataset, we thank the NIH Clinical Center for making it publicly available. For tuberculosis data collection, thanks go to Sameer Antani, Stefan Jaeger, Sema Candemir, Zhiyun Xue, Alex Karargyris, George R. Thomas, Pu-Xuan Lu, Yi-Xiang Wang, Michael Bonifant, Ellan Kim, Sonia Qasba, and Jonathan Musco. The authors would also like to acknowledge many members of the Google Health Radiology and labeling software teams, in particular Shruthi Prabhakara, Scott McKinney, and Akib Uddin. Sincere appreciation also goes to the radiologists who enabled this work with their image interpretation and annotation efforts throughout the study; Jonny Wong for coordinating the imaging annotation work; Gavin Bee, Mikhail Fomitchev, Shabir Adeel, Jeff Bertram, and Benedict Noero for data releasing; David F. Steiner, Kunal Nagpal, and Michael D. Howell for providing feedback on the manuscript; Craig Mermel, Lauren Winer, Johnny Luu, Adrienne Welch, Annisah Um’rani, and Ashley Zlatinov for feedback on the blogpost.

---

¹Labels include atelectasis, cardiomegaly, effusion, infiltration, mass, nodule, pneumonia, pneumothorax, consolidation, edema, emphysema, fibrosis, pleural thickening, hernia, other abnormality, and normal vs abnormal. ^↩

Posted by Forrester Cole, Software Engineer and Tali Dekel, Research Scientist

Image and video editing operations often rely on accurate mattes — images that define a separation between foreground and background. While recent computer vision techniques can produce high-quality mattes for natural images and videos, allowing real-world applications such as generating synthetic depth-of-field, editing and synthesising images, or removing backgrounds from images, one fundamental piece is missing: the various scene effects that the subject may generate, like shadows, reflections, or smoke, are typically overlooked.

In “Omnimatte: Associating Objects and Their Effects in Video”, presented at CVPR 2021, we describe a new approach to matte generation that leverages layered neural rendering to separate a video into layers called omnimattes that include not only the subjects but also all of the effects related to them in the scene. Whereas a typical state-of-the-art segmentation model extracts masks for the subjects in a scene, for example, a person and a dog, the method proposed here can isolate and extract additional details associated with the subjects, such as shadows cast on the ground.

A state-of-the-art segmentation network (e.g., MaskRCNN) takes an input video (left) and produces plausible masks for people and animals (middle), but misses their associated effects. Our method produces mattes that include not only the subjects, but their shadows as well (right; individual channels for person and dog visualized as blue and green).

Also unlike segmentation masks, omnimattes can capture partially-transparent, soft effects such as reflections, splashes, or tire smoke. Like conventional mattes, omnimattes are RGBA images that can be manipulated using widely-available image or video editing tools, and can be used wherever conventional mattes are used, for example, to insert text into a video underneath a smoke trail.

Layered Decomposition of Video
To generate omnimattes, we split the input video into a set of layers: one for each moving subject, and one additional layer for stationary background objects. In the example below, there is one layer for the person, one for the dog, and one for the background. When merged together using conventional alpha blending, these layers reproduce the input video.

Besides reproducing the video, the decomposition must capture the correct effects in each layer. For example, if the person’s shadow appears in the dog’s layer, the merged layers would still reproduce the input video, but inserting an additional element between the person and dog would produce an obvious error. The challenge is to find a decomposition where each subject’s layer captures only that subject’s effects, producing a true omnimatte.

Our solution is to apply our previously developed layered neural rendering approach to train a convolutional neural network (CNN) to map the subject’s segmentation mask and a background noise image into an omnimatte. Due to their structure, CNNs are naturally inclined to learn correlations between image effects, and the stronger the correlation between the effects, the easier for the CNN to learn. In the above video, for example, the spatial relationships between the person and their shadow, and the dog and its shadow, remain similar as they walk from right to left. The relationships change more (hence, the correlations are weaker) between the person and the dog’s shadow, or the dog and the person’s shadow. The CNN learns the stronger correlations first, leading to the correct decomposition.

The omnimatte system is shown in detail below. In a preprocess, the user chooses the subjects and specifies a layer for each. A segmentation mask for each subject is extracted using an off-the-shelf segmentation network, such as MaskRCNN, and camera transformations relative to the background are found using standard camera stabilization tools. A random noise image is defined in the background reference frame and sampled using the camera transformations to produce per-frame noise images. The noise images provide image features that are random but consistently track the background over time, providing a natural input for the CNN to learn to reconstruct the background colors.

The rendering CNN takes as input the segmentation mask and the per-frame noise images and produces the RGB color images and alpha maps, which capture the transparency of each layer. These outputs are merged using conventional alpha-blending to produce the output frame. The CNN is trained from scratch to reconstruct the input frames by finding and associating the effects not captured in a mask (e.g., shadows, reflections or smoke) with the given foreground layer, and to ensure the subject’s alpha roughly includes the segmentation mask. To make sure the foreground layers only capture the foreground elements and none of the stationary background, a sparsity loss is also applied on the foreground alpha.

A new rendering network is trained for each video. Because the network is only required to reconstruct the single input video, it is able to capture fine structures and fast motion in addition to separating the effects of each subject, as seen below. In the walking example, the omnimatte includes the shadow cast on the slats of the park bench. In the tennis example, the thin shadow and even the tennis ball are captured. In the soccer example, the shadow of the player and the ball are decomposed into their proper layers (with a slight error when the player’s foot is occluded by the ball).

This basic model already works well, but one can improve the results by augmenting the input of the CNN with additional buffers such as optical flow or texture coordinates.

Applications
Once the omnimattes are generated, how can they be used? As shown above, we can remove objects, simply by removing their layer from the composition. We can also duplicate objects, by repeating their layer in the composition. In the example below, the video has been “unwrapped” into a panorama, and the horse duplicated several times to produce a stroboscopic photograph effect. Note that the shadow that the horse casts on the ground and onto the obstacle is correctly captured.

A more subtle, but powerful application is to retime the subjects. Manipulation of time is widely used in film, but usually requires separate shots for each subject and a controlled filming environment. A decomposition into omnimattes makes retiming effects possible for everyday videos using only post-processing, simply by independently changing the playback rate of each layer. Since the omnimattes are standard RGBA images, this retiming edit can be done using conventional video editing software.

The video below is decomposed into three layers, one for each child. The children’s initial, unsynchronized jumps are aligned by simply adjusting the playback rate of their layers, producing realistic retiming for the splashes and reflections in the water.

In the original video (left), each child jumps at a different time. After editing (right), everyone jumps together.

It’s important to consider that any novel technique for manipulating images should be developed and applied responsibly, as it could be misused to produce fake or misleading information. Our technique was developed in accordance with our AI Principles and only allows rearrangement of content already present in the video, but even simple rearrangement can significantly alter the effect of a video, as shown in these examples. Researchers should be aware of these risks.

Future Work
There are a number of exciting directions to improve the quality of the omnimattes. On a practical level, this system currently only supports backgrounds that can be modeled as panoramas, where the position of the camera is fixed. When the camera position moves, the panorama model cannot accurately capture the entire background, and some background elements may clutter the foreground layers (sometimes visible in the above figures). Handling fully general camera motion, such as walking through a room or down a street, would require a 3D background model. Reconstruction of 3D scenes in the presence of moving objects and effects is still a difficult research challenge, but one that has seen promising recent progress.

On a theoretical level, the ability of CNNs to learn correlations is powerful, but still somewhat mysterious, and does not always lead to the expected layer decomposition. While our system allows for manual editing when the automatic result is imperfect, a better solution would be to fully understand the capabilities and limitations of CNNs to learn image correlations. Such an understanding could lead to improved denoising, inpainting, and many other video editing applications besides layer decomposition.

Acknowledgements
Erika Lu, from the University of Oxford, developed the omnimatte system during two internships at Google, in collaboration with Google researchers Forrester Cole, Tali Dekel, Michael Rubinstein, William T. Freeman and David Salesin, and University of Oxford researchers Weidi Xie and Andrew Zisserman.

Thank you to the friends and families of the authors who agreed to appear in the example videos. The “horse jump low”, “lucia”, and “tennis” videos are from the DAVIS 2016 dataset. The soccer video is used by permission from Online Soccer Skills. The car drift video was licensed from Shutterstock.

Posted by Ye Jia, Software Engineer and Julie Cattiau, Product Manager, Google Research

On June 2nd, 2021, Major League Baseball in the United States celebrated Lou Gehrig Day, commemorating both the day in 1925 that Lou Gehrig became the Yankees’ starting first baseman, and the day in 1941 that he passed away from amyotrophic lateral sclerosis (ALS, also known as Lou Gehrig’s disease) at the age of 37. ALS is a progressive neurodegenerative disease that affects motor neurons, which connect the brain with the muscles throughout the body, and govern muscle control and voluntary movements. When voluntary muscle control is affected, people may lose their ability to speak, eat, move and breathe.

In honor of Lou Gehrig, former NFL player and ALS advocate Steve Gleason, who lost his ability to speak due to ALS, recited Gehrig’s famous “Luckiest Man” speech at the June 2nd event using a recreation of his voice generated by a machine learning (ML) model. Gleason’s voice recreation was developed in collaboration with Google’s Project Euphonia, which aims to empower people who have impaired speaking ability due to ALS to better communicate using their own voices.

Steve Gleason, who lost his voice to ALS, worked with Google’s Project Euphonia to generate a speech in his own voice in honor of Lou Gehrig. A portion of Gleason’s speech was broadcast in ballparks across the country during the 4th inning on June 2nd, 2021.

Today we describe PnG NAT, the model adopted by Project Euphonia to recreate Steve Gleason’s voice. PnG NAT is a new text-to-speech synthesis (TTS) model that merges two state-of-the-art technologies, PnG BERT and Non-Attentive Tacotron (NAT), into a single model. It demonstrates significantly better quality and fluency than previous technologies, and represents a promising approach that can be extended to a wider array of users.

Recreating a Voice
Non-Attentive Tacotron (NAT) is the successor to Tacotron 2, a sequence-to-sequence neural TTS model proposed in 2017. Tacotron 2 used an attention module to connect the input text sequence and the output speech spectrogram frame sequence, so that the model knows which part of the text to pay attention to when generating each time step of the synthesized speech spectrogram. Tacotron 2 was the first TTS model that was able to synthesize speech that sounds as natural as a person speaking. However, with extensive experimentation we discovered that there is a small probability that the model can suffer from robustness issues — such as babbling, repeating, or skipping part of the text — due to the inherent flexibility of the attention mechanism.

NAT improves upon Tacotron 2 by replacing the attention module with a duration-based upsampler, which predicts a duration for each input phoneme and upsamples the encoded phoneme representation so that the output length corresponds to the length of the predicted speech spectrogram. Such a change both resolves the robustness issue, and improves the naturalness of the synthesized speech. This approach also enables precise control of the speech duration for each phoneme of the input text while still maintaining highly natural synthesis quality. Because recordings of people with ALS often exhibit disfluent speech, this ability to exert per-phoneme control is key for achieving the fluency of the recreated voice.

Non-Attentive Tacotron (NAT) model.

While NAT addresses the robustness issue and enables precise duration control in neural TTS, we build upon it to further improve the natural language understanding of the TTS input. For this, we apply PnG BERT, which uses an approach similar to BERT, but is specifically designed for TTS. It is pre-trained with self-supervision on both the phoneme representation and the grapheme representation of the same content from a large text corpus, and then is used as the encoder of the TTS model. This results in a significant improvement of the prosody and pronunciation of the synthesized speech, especially in difficult cases.

Take, for example, the following audio, which was synthesized from a regular NAT model that takes only phonemes as input:

In comparison, the audio synthesized from PnG NAT on the same input text includes an additional pause that makes the meaning more clear.

The input text to both models is, “To cancel the payment, press one; or to continue, two.” Notice the different pause lengths before the ending “two” in the two versions. The word “two” in the version output by the regular NAT model could be confused for “too”. Because “too” and “two” have identical pronunciation (and thus the same phoneme representation), the regular NAT model does not understand which of the two is appropriate, and assumes it to be the word that more frequently follows a comma, “too”. In contrast, the PnG NAT model can more easily tell the difference, because it takes graphemes in addition to phonemes as input, and thus makes more appropriate pause.

The PnG NAT model integrates the pre-trained PnG BERT model as the encoder to the NAT model. The hidden representations output from the encoder are used by NAT to predict the duration of each phoneme, and are then upsampled to match the length of the audio spectrogram, as outlined above. In the final step, a non-attentive decoder converts the upsampled hidden representations into audio speech spectrograms, which are finally converted into audio waveforms by a neural vocoder.

PnG BERT and the pre-training objectives. Yellow boxes represent phonemes, and pink boxes represent graphemes.

PnG NAT: PnG BERT replaces the original encoder in the NAT model. The random masking for the Masked Language Model (MLM) pre-training is removed.

To recreate Steve Gleason’s voice, we first trained a PnG NAT model with recordings from 31 professional speakers, and then fine-tuned it with 30 minutes of Gleason’s recordings. Because these latter recordings were made after he was diagnosed with ALS, they exhibit signs of slurring. The fine tuned model was able to synthesize speech that sounds very similar to these recordings. However, because the symptoms of ALS were already present in Gleason’s speech, they exhibited some similar disfluencies.

To mitigate this, we leveraged the phoneme duration control of NAT as well as the model trained with professional speakers. We first predicted the durations of each phoneme for both a professional speaker and for Gleason, and then used the geometric mean of the two durations for each phoneme to guide the NAT output. As a result, the model is able to speak in Gleason’s voice, but more fluently than in the original recordings.

Here is the full version of the synthesized Lou Gehrig speech in Gleason’s voice:

<!– As a comparison, following is one of Gleason’s recordings that was used to train the model:

–>

Besides recreating voices for people with ALS, PnG NAT is also powering voices for a variety of customers through Google Cloud Custom Voice.

Project Euphonia
Of the millions of people around the world who have neurologic conditions that may impact their speech, such as ALS, cerebral palsy or Down syndrome, many may find it difficult to be understood, which can make face-to-face communication challenging. Using voice-activated technologies can be frustrating too, as they don’t always work reliably. Project Euphonia is a Google Research initiative focused on helping people with impaired speech be better understood. The team is researching ways to improve speech recognition for individuals with speech impairments (see recent blog post and segment in TODAY show), as well as customized text-to-speech technology (see Age of AI documentary featuring former NFL player Tim Shaw).

Acknowledgements
Many people across Google Research, Google Cloud and Consumer Apps, and Google Accessibility teams contributed to this project and the event, including Michael Brenner, Bob MacDonald, Heiga Zen, Yu Zhang, Jonathan Shen, Isaac Elias‎, Yonghui Wu, Anne Keck, Danielle Notaro, Kevin Hogan, Zack Kaplan, KR Liu, Kyndra Price, Zoe Ortiz.

Geometric deep learning is a “program” that aspires to situate deep learning architectures and techniques in a framework of mathematical priors. The priors, such as various types of invariance, first arise in some physical domain. A neural network that well matches the domain will preserve as many invariances as possible. In this post, we present a very conceptual, high-level overview, and highlight a few applications.

Posted by Neil Zeghidour, Research Scientist and Marco Tagliasacchi, Staff Research Scientist, Google Research

Audio codecs are used to efficiently compress audio to reduce either storage requirements or network bandwidth. Ideally, audio codecs should be transparent to the end user, so that the decoded audio is perceptually indistinguishable from the original and the encoding/decoding process does not introduce perceivable latency.

Over the past few years, different audio codecs have been successfully developed to meet these requirements, including Opus and Enhanced Voice Services (EVS). Opus is a versatile speech and audio codec, supporting bitrates from 6 kbps (kilobits per second) to 510 kbps, which has been widely deployed across applications ranging from video conferencing platforms, like Google Meet, to streaming services, like YouTube. EVS is the latest codec developed by the 3GPP standardization body targeting mobile telephony. Like Opus, it is a versatile codec operating at multiple bitrates, 5.9 kbps to 128 kbps. The quality of the reconstructed audio using either of these codecs is excellent at medium-to-low bitrates (12–20 kbps), but it degrades sharply when operating at very low bitrates (⪅3 kbps). While these codecs leverage expert knowledge of human perception as well as carefully engineered signal processing pipelines to maximize the efficiency of the compression algorithms, there has been recent interest in replacing these handcrafted pipelines by machine learning approaches that learn to encode audio in a data-driven manner.

Earlier this year, we released Lyra, a neural audio codec for low-bitrate speech. In “SoundStream: an End-to-End Neural Audio Codec”, we introduce a novel neural audio codec that extends those efforts by providing higher-quality audio and expanding to encode different sound types, including clean speech, noisy and reverberant speech, music, and environmental sounds. SoundStream is the first neural network codec to work on speech and music, while being able to run in real-time on a smartphone CPU. It is able to deliver state-of-the-art quality over a broad range of bitrates with a single trained model, which represents a significant advance in learnable codecs.

Learning an Audio Codec from Data
The main technical ingredient of SoundStream is a neural network, consisting of an encoder, decoder and quantizer, all of which are trained end-to-end. The encoder converts the input audio stream into a coded signal, which is compressed using the quantizer and then converted back to audio using the decoder. SoundStream leverages state-of-the-art solutions in the field of neural audio synthesis to deliver audio at high perceptual quality, by training a discriminator that computes a combination of adversarial and reconstruction loss functions that induce the reconstructed audio to sound like the uncompressed original input. Once trained, the encoder and decoder can be run on separate clients to efficiently transmit high-quality audio over a network.

SoundStream training and inference. During training, the encoder, quantizer and decoder parameters are optimized using a combination of reconstruction and adversarial losses, computed by a discriminator, which is trained to distinguish between the original input audio and the reconstructed audio. During inference, the encoder and quantizer on a transmitter client send the compressed bitstream to a receiver client that can then decode the audio signal.

Learning a Scalable Codec with Residual Vector Quantization
The encoder of SoundStream produces vectors that can take an indefinite number of values. In order to transmit them to the receiver using a limited number of bits, it is necessary to replace them by close vectors from a finite set (called a codebook), a process known as vector quantization. This approach works well at bitrates around 1 kbps or lower, but quickly reaches its limits when using higher bitrates. For example, even at a bitrate as low as 3 kbps, and assuming the encoder produces 100 vectors per second, one would need to store a codebook with more than 1 billion vectors, which is infeasible in practice.

In SoundStream, we address this issue by proposing a new residual vector quantizer (RVQ), consisting of several layers (up to 80 in our experiments). The first layer quantizes the code vectors with moderate resolution, and each of the following layers processes the residual error from the previous one. By splitting the quantization process in several layers, the codebook size can be reduced drastically. As an example, with 100 vectors per second at 3 kbps, and using 5 quantizer layers, the codebook size goes from 1 billion to 320. Moreover, we can easily increase or decrease the bitrate by adding or removing quantizer layers, respectively.

Because network conditions can vary while transmitting audio, ideally a codec should be “scalable” so that it can change its bitrate from low to high depending on the state of the network. While most traditional codecs are scalable, previous learnable codecs need to be trained and deployed specifically for each bitrate.

To circumvent this limitation, we leverage the fact that the number of quantization layers in SoundStream controls the bitrate, and propose a new method called “quantizer dropout”. During training, we randomly drop some quantization layers to simulate a varying bitrate. This pushes the decoder to perform well at any bitrate of the incoming audio stream, and thus helps SoundStream to become “scalable” so that a single trained model can operate at any bitrate, performing as well as models trained specifically for these bitrates.

Comparison of SoundStream models (higher is better) that are trained at 18 kbps with quantizer dropout (bitrate scalable), without quantizer dropout (not bitrate scalable) and evaluated with a variable number of quantizers, or trained and evaluated at a fixed bitrate (bitrate specific). The bitrate-scalable model (a single model for all bitrates) does not lose any quality when compared to bitrate-specific models (a different model for each bitrate), thanks to quantizer dropout.

A State-of-the-Art Audio Codec
SoundStream at 3 kbps outperforms Opus at 12 kbps and approaches the quality of EVS at 9.6 kbps, while using 3.2x–4x fewer bits. This means that encoding audio with SoundStream can provide a similar quality while using a significantly lower amount of bandwidth. Moreover, at the same bitrate, SoundStream outperforms the current version of Lyra, which is based on an autoregressive network. Unlike Lyra, which is already deployed and optimized for production usage, SoundStream is still at an experimental stage. In the future, Lyra will incorporate the components of SoundStream to provide both higher audio quality and reduced complexity.

SoundStream at 3kbps vs. state-of-the-art codecs. MUSHRA score is an indication of subjective quality (the higher the better).

The demonstration of SoundStream’s performance compared to Opus, EVS, and the original Lyra codec is presented in these audio examples, a selection of which are provided below.

Speech

Reference
Lyra (3kbps)
Opus (6kbps)
EVS (5.9kbps)
SoundStream (3kbps)

Music

Reference
Lyra (3kbps)
Opus (6kbps)
EVS (5.9kbps)
SoundStream (3kbps)

Joint Audio Compression and Enhancement
In traditional audio processing pipelines, compression and enhancement (the removal of background noise) are typically performed by different modules. For example, it is possible to apply an audio enhancement algorithm at the transmitter side, before audio is compressed, or at the receiver side, after audio is decoded. In such a setup, each processing step contributes to the end-to-end latency. Conversely, we design SoundStream in such a way that compression and enhancement can be carried out jointly by the same model, without increasing the overall latency. In the following examples, we show that it is possible to combine compression with background noise suppression, by activating and deactivating denoising dynamically (no denoising for 5 seconds, denoising for 5 seconds, no denoising for 5 seconds, etc.).

Original noisy audio
Denoised output*

* Demonstrated by turning denoising on and off every 5 seconds.

Conclusion
Efficient compression is necessary whenever one needs to transmit audio, whether when streaming a video, or during a conference call. SoundStream is an important step towards improving machine learning-driven audio codecs. It outperforms state-of-the-art codecs, such as Opus and EVS, can enhance audio on demand, and requires deployment of only a single scalable model, rather than many.

SoundStream will be released as a part of the next, improved version of Lyra. By integrating SoundStream with Lyra, developers can leverage the existing Lyra APIs and tools for their work, providing both flexibility and better sound quality. We will also release it as a separate TensorFlow model for experimentation.

AcknowledgmentsThe work described here was authored by Neil Zeghidour, Alejandro Luebs, Ahmed Omran, Jan Skoglund and Marco Tagliasacchi. We are grateful for all discussions and feedback on this work that we received from our colleagues at Google.