Posts in category: Google AI

Posted by AJ Piergiovanni and Anelia Angelova, Research Scientists, Google Research, Brain Team

Video is an ubiquitous source of media content that touches on many aspects of people’s day-to-day lives. Increasingly, real-world video applications, such as video captioning, video content analysis, and video question-answering (VideoQA), rely on models that can connect video content with text or natural language. VideoQA is particularly challenging, however, as it requires grasping both semantic information, such as objects in a scene, as well as temporal information, e.g., how things move and interact, both of which must be taken in the context of a natural-language question that holds specific intent. In addition, because videos have many frames, processing all of them to learn spatio-temporal information can be computationally expensive. Nonetheless, understanding all this information enables models to answer complex questions — for example, in the video below, a question about the second ingredient poured in the bowl requires identifying objects (the ingredients), actions (pouring), and temporal ordering (second).

An example input question for the VideoQA task “What is the second ingredient poured into the bowl?” which requires deeper understanding of both the visual and text inputs. The video is an example from the 50 Salads dataset, used under the Creative Commons license.

To address this, in “Video Question Answering with Iterative Video-Text Co-Tokenization”, we introduce a new approach to video-text learning called iterative co-tokenization, which is able to efficiently fuse spatial, temporal and language information for VideoQA. This approach is multi-stream, processing different scale videos with independent backbone models for each to produce video representations that capture different features, e.g., those of high spatial resolution or long temporal durations. The model then applies the co-tokenization module to learn efficient representations from fusing the video streams with the text. This model is highly efficient, using only 67 giga-FLOPs (GFLOPs), which is at least 50% fewer than previous approaches, while giving better performance than alternative state-of-the-art models.

Video-Text Iterative Co-tokenization
The main goal of the model is to produce features from both videos and text (i.e., the user question), jointly allowing their corresponding inputs to interact. A second goal is to do so in an efficient manner, which is highly important for videos since they contain tens to hundreds of frames as input.

The model learns to tokenize the joint video-language inputs into a smaller set of tokens that jointly and efficiently represent both modalities. When tokenizing, we use both modalities to produce a joint compact representation, which is fed to a transformer layer to produce the next level representation. A challenge here, which is also typical in cross-modal learning, is that often the video frame does not correspond directly to the associated text. We address this by adding two learnable linear layers which unify the visual and text feature dimensions before tokenization. This way we enable both video and text to condition how video tokens are learned.

Moreover, a single tokenization step does not allow for further interaction between the two modalities. For that, we use this new feature representation to interact with the video input features and produce another set of tokenized features, which are then fed into the next transformer layer. This iterative process allows the creation of new features, or tokens, which represent a continual refinement of the joint representation from both modalities. At the last step the features are input to a decoder that generates the text output.

As customarily done for VideoQA, we pre-train the model before fine-tuning it on the individual VideoQA datasets. In this work we use the videos automatically annotated with text based on speech recognition, using the HowTo100M dataset instead of pre-training on a large VideoQA dataset. This weaker pre-training data still enables our model to learn video-text features.

Visualization of the video-text iterative co-tokenization approach. Multi-stream video inputs, which are versions of the same video input (e.g., a high resolution, low frame-rate video and a low resolution, high frame-rate video), are efficiently fused together with the text input to produce a text-based answer by the decoder. Instead of processing the inputs directly, the video-text iterative co-tokenization model learns a reduced number of useful tokens from the fused video-language inputs. This process is done iteratively, allowing the current feature tokenization to affect the selection of tokens at the next iteration, thus refining the selection.

Efficient Video Question-Answering
We apply the video-language iterative co-tokenization algorithm to three main VideoQA benchmarks, MSRVTT-QA, MSVD-QA and IVQA, and demonstrate that this approach achieves better results than other state-of-the-art models, while having a modest size. Furthermore, iterative co-tokenization learning yields significant compute savings for video-text learning tasks. The method uses only 67 giga-FLOPs (GFLOPS), which is one sixth the 360 GFLOPS needed when using the popular 3D-ResNet video model jointly with text and is more than twice as efficient as the X3D model. This is all the while producing highly accurate results, outperforming state-of-the-art methods.

Comparison of our iterative co-tokenization approach to previous methods such as MERLOT and VQA-T, as well as, baselines using single ResNet-3D or X3D-XL.

Multi-stream Video Inputs
For VideoQA, or any of a number of other tasks that involve video inputs, we find that multi-stream input is important to more accurately answer questions about both spatial and temporal relationships. Our approach utilizes three video streams at different resolutions and frame-rates: a low-resolution high frame-rate, input video stream (with 32 frames-per-second and spatial resolution 64×64, which we denote as 32x64x64); a high-resolution, low frame-rate video (8x224x224); and one in-between (16x112x112). Despite the apparently more voluminous information to process with three streams, we obtain very efficient models due to the iterative co-tokenization approach. At the same time these additional streams allow extraction of the most pertinent information. For example, as shown in the figure below, questions related to a specific activity in time will produce higher activations in the smaller resolution but high frame-rate video input, whereas questions related to the general activity can be answered from the high resolution input with very few frames. Another benefit of this algorithm is that the tokenization changes depending on the questions asked.

Visualization of the attention maps learned per layer during the video-text co-tokenization. The attention maps differ depending on the question asked for the same video. For example, if the question is related to the general activity (e.g., surfing in the figure above), then the attention maps of the higher resolution low frame-rate inputs are more active and seem to consider more global information. Whereas if the question is more specific, e.g., asking about what happens after an event, the feature maps are more localized and tend to be active in the high frame-rate video input. Furthermore, we see that the low-resolution, high-frame rate video inputs provide more information related to activities in the video.

Conclusion
We present a new approach to video-language learning that focuses on joint learning across video-text modalities. We address the important and challenging task of video question-answering. Our approach is both highly efficient and accurate, outperforming current state-of-the-art models, despite being more efficient. Our approach results in modest model sizes and can gain further improvements with larger models and data. We hope this work provokes more research in vision-language learning to enable more seamless interaction with vision-based media.

Acknowledgements
This work is conducted by AJ Pierviovanni, Kairo Morton, Weicheng Kuo, Michael Ryoo and Anelia Angelova. We thank our collaborators in this research, and Soravit Changpinyo for valuable comments and suggestions, and Claire Cui for suggestions and support. We also thank Tom Small for visualizations.

Read More

Posted by Bingyi Cao, Software Engineer, Google Research, and Mário Lipovský, Software Engineer, Google Lens

Computer vision models see daily application for a wide variety of tasks, ranging from object recognition to image-based 3D object reconstruction. One challenging type of computer vision problem is instance-level recognition (ILR) — given an image of an object, the task is to not only determine the generic category of an object (e.g., an arch), but also the specific instance of the object (”Arc de Triomphe de l’Étoile, Paris, France”).

Previously, ILR was tackled using deep learning approaches. First, a large set of images was collected. Then a deep model was trained to embed each image into a high-dimensional space where similar images have similar representations. Finally, the representation was used to solve the ILR tasks related to classification (e.g., with a shallow classifier trained on top of the embedding) or retrieval (e.g., with a nearest neighbor search in the embedding space).

Since there are many different object domains in the world, e.g., landmarks, products, or artworks, capturing all of them in a single dataset and training a model that can distinguish between them is quite a challenging task. To decrease the complexity of the problem to a manageable level, the focus of research so far has been to solve ILR for a single domain at a time. To advance the research in this area, we hosted multiple Kaggle competitions focused on the recognition and retrieval of landmark images. In 2020, Amazon joined the effort and we moved beyond the landmark domain and expanded to the domains of artwork and product instance recognition. The next step is to generalize the ILR task to multiple domains.

To this end, we’re excited to announce the Google Universal Image Embedding Challenge, hosted by Kaggle in collaboration with Google Research and Google Lens. In this challenge, we ask participants to build a single universal image embedding model capable of representing objects from multiple domains at the instance level. We believe that this is the key for real-world visual search applications, such as augmenting cultural exhibits in a museum, organizing photo collections, visual commerce and more.

Images1 of object instances from some domains represented in the dataset: apparel and accessories, furniture and home goods, toys, cars, landmarks, dishes, artwork and illustrations.

Degrees of Variation in Different Domains
To represent objects from a large number of domains, we require one model to learn many domain-specific subtasks (e.g., filtering different kinds of noise or focusing on a specific detail), which can only be learned from a semantically and visually diverse collection of images. Addressing each degree of variation proposes a new challenge for both image collection and model training.

The first sort of variation comes from the fact that while some domains contain unique objects in the world (landmarks, artwork, etc.), others contain objects that may have many copies (clothing, furniture, packaged goods, food, etc.). Because a landmark is always placed at the same location, the surrounding context may be useful for recognition. In contrast, a product, say a phone, even of a specific model and color, may have millions of physical instances and thus appear in many surrounding contexts.

Another challenge comes from the fact that a single object may appear different depending on the point of view, lighting conditions, occlusion or deformations (e.g., a dress worn on a person may look very different than on a hanger). In order for a model to learn invariance to all of these visual modes, all of them should be captured by the training data.

Additionally, similarities between objects differ across domains. For example, in order for a representation to be useful in the product domain, it must be able to distinguish very fine-grained details between similarly looking products belonging to two different brands. In the domain of food, however, the same dish (e.g., spaghetti bolognese) cooked by two chefs may look quite different, but the ability of the model to distinguish spaghetti bolognese from other dishes may be sufficient for the model to be useful. Additionally, a vision model of high quality should assign similar representations to more visually similar renditions of a dish.

<!–

–><!–

–>

Domain			Landmark			Apparel
Image
Instance Name			Empire State Building2			Cycling jerseys with Android logo3
Which physical objects belong to the instance class?			Single instance in the world			Many physical instances; may differ in size or pattern (e.g., a patterned cloth cut differently)
What are the possible views of the object?			Appearance variation only based on capture conditions (e.g., illumination or viewpoint); limited number of common external views; possibility of many internal views			Deformable appearance (e.g., worn or not); limited number of common views: front, back, side
What are the surroundings and are they useful for recognition?			Surrounding context does not vary much other than daily and yearly cycles; may be useful for verifying the object of interest			Surrounding context can change dramatically due to difference in environment, additional pieces of clothing, or accessories partially occluding clothing of interest (e.g., a jacket or a scarf)
What may be tricky cases that do not belong to the instance class?			Replicas of landmarks (e.g., Eiffel Tower in Las Vegas), souvenirs			Same piece of apparel of different material or different color; visually very similar pieces with a small distinguishing detail (e.g., a small brand logo); different pieces of apparel worn by the same model

Learning Multi-domain Representations
After a collection of images covering a variety of domains is created, the next challenge is to train a single, universal model. Some features and tasks, such as representing color, are useful across many domains, and thus adding training data from any domain will likely help the model improve at distinguishing colors. Other features may be more specific to selected domains, thus adding more training data from other domains may deteriorate the model’s performance. For example, while for 2D artwork it may be very useful for the model to learn to find near duplicates, this may deteriorate the performance on clothing, where deformed and occluded instances need to be recognized.

The large variety of possible input objects and tasks that need to be learned require novel approaches for selecting, augmenting, cleaning and weighing the training data. New approaches for model training and tuning, and even novel architectures may be required.

Universal Image Embedding Challenge
To help motivate the research community to address these challenges, we are hosting the Google Universal Image Embedding Challenge. The challenge was launched on Kaggle in July and will be open until October, with cash prizes totaling $50k. The winning teams will be invited to present their methods at the Instance-Level Recognition workshop at ECCV 2022.

Participants will be evaluated on a retrieval task on a dataset of ~5,000 test query images and ~200,000 index images, from which similar images are retrieved. In contrast to ImageNet, which includes categorical labels, the images in this dataset are labeled at the instance level.

The evaluation data for the challenge is composed of images from the following domains: apparel and accessories, packaged goods, furniture and home goods, toys, cars, landmarks, storefronts, dishes, artwork, memes and illustrations.

Distribution of domains of query images.

We invite researchers and machine learning enthusiasts to participate in the Google Universal Image Embedding Challenge and join the Instance-Level Recognition workshop at ECCV 2022. We hope the challenge and the workshop will advance state-of-the-art techniques on multi-domain representations.

Acknowledgement
The core contributors to this project are Andre Araujo, Boris Bluntschli, Bingyi Cao, Kaifeng Chen, Mário Lipovský, Grzegorz Makosa, Mojtaba Seyedhosseini and Pelin Dogan Schönberger. We would like to thank Sohier Dane, Will Cukierski and Maggie Demkin for their help organizing the Kaggle challenge, as well as our ECCV workshop co-organizers Tobias Weyand, Bohyung Han, Shih-Fu Chang, Ondrej Chum, Torsten Sattler, Giorgos Tolias, Xu Zhang, Noa Garcia, Guangxing Han, Pradeep Natarajan and Sanqiang Zhao. Furthermore we are thankful to Igor Bonaci, Tom Duerig, Vittorio Ferrari, Victor Gomes, Futang Peng and Howard Zhou who gave us feedback, ideas and support at various points of this project.

¹ Image credits: Chris Schrier, CC-BY; Petri Krohn, GNU Free Documentation License; Drazen Nesic, CC0; Marco Verch Professional Photographer, CCBY; Grendelkhan, CCBY; Bobby Mikul, CC0; Vincent Van Gogh, CC0; pxhere.com, CC0; Smart Home Perfected, CC-BY.  ^↩
2 Image credit: Bobby Mikul, CC0.  ^↩
3 Image credit: Chris Schrier, CC-BY.  ^↩

Read More

Posted by Qifei Wang, Senior Software Engineer, and Feng Yang, Senior Staff Software Engineer, Google Research

Deep learning models for visual tasks (e.g., image classification) are usually trained end-to-end with data from a single visual domain (e.g., natural images or computer generated images). Typically, an application that completes visual tasks for multiple domains would need to build multiple models for each individual domain, train them independently (meaning no data is shared between domains), and then at inference time each model would process domain-specific input data. However, early layers between these models generate similar features, even for different domains, so it can be more efficient — decreasing latency and power consumption, lower memory overhead to store parameters of each model — to jointly train multiple domains, an approach referred to as multi-domain learning (MDL). Moreover, an MDL model can also outperform single domain models due to positive knowledge transfer, which is when additional training on one domain actually improves performance for another. The opposite, negative knowledge transfer, can also occur, depending on the approach and specific combination of domains involved. While previous work on MDL has proven the effectiveness of jointly learning tasks across multiple domains, it involved a hand-crafted model architecture that is inefficient to apply to other work.

In “Multi-path Neural Networks for On-device Multi-domain Visual Classification”, we propose a general MDL model that can: 1) achieve high accuracy efficiently (keeping the number of parameters and FLOPS low), 2) learn to enhance positive knowledge transfer while mitigating negative transfer, and 3) effectively optimize the joint model while handling various domain-specific difficulties. As such, we propose a multi-path neural architecture search (MPNAS) approach to build a unified model with heterogeneous network architecture for multiple domains. MPNAS extends the efficient neural architecture search (NAS) approach from single path search to multi-path search by finding an optimal path for each domain jointly. Also, we introduce a new loss function, called adaptive balanced domain prioritization (ABDP) that adapts to domain-specific difficulties to help train the model efficiently. The resulting MPNAS approach is efficient and scalable; the resulting model maintains performance while reducing the model size and FLOPS by 78% and 32%, respectively, compared to a single-domain approach.

Multi-Path Neural Architecture Search
To encourage positive knowledge transfer and avoid negative transfer, traditional solutions build an MDL model so that domains share most of the layers that learn the shared features across domains (called feature extraction), then have a few domain-specific layers on top. However, such a homogenous approach to feature extraction cannot handle domains with significantly different features (e.g., objects in natural images and art paintings). On the other hand, handcrafting a unified heterogeneous architecture for each MDL model is time-consuming and requires domain-specific knowledge.

NAS is a powerful paradigm for automatically designing deep learning architectures. It defines a search space, made up of various potential building blocks that could be part of the final model. The search algorithm finds the best candidate architecture from the search space that optimizes the model objectives, e.g., classification accuracy. Recent NAS approaches (e.g., TuNAS) have meaningfully improved search efficiency by using end-to-end path sampling, which enables us to scale NAS from single domains to MDL.

Inspired by TuNAS, MPNAS builds the MDL model architecture in two stages: search and training. In the search stage, to find an optimal path for each domain jointly, MPNAS creates an individual reinforcement learning (RL) controller for each domain, which samples an end-to-end path (from input layer to output layer) from the supernetwork (i.e., the superset of all the possible subnetworks between the candidate nodes defined by the search space). Over multiple iterations, all the RL controllers update the path to optimize the RL rewards across all domains. At the end of the search stage, we obtain a subnetwork for each domain. Finally, all the subnetworks are combined to build a heterogeneous architecture for the MDL model, shown below.

Since the subnetwork for each domain is searched independently, the building block in each layer can be shared by multiple domains (i.e., dark gray nodes), used by a single domain (i.e., light gray nodes), or not used by any subnetwork (i.e., dotted nodes). The path for each domain can also skip any layer during search. Given the subnetwork can freely select which blocks to use along the path in a way that optimizes performance (rather than, e.g., arbitrarily designating which layers are homogenous and which are domain-specific), the output network is both heterogeneous and efficient.

Example architecture searched by MPNAS. Dashed paths represent all the possible subnetworks. Solid paths represent the selected subnetworks for each domain (highlighted in different colors). Nodes in each layer represent the candidate building blocks defined by the search space.

The figure below demonstrates the searched architecture of two visual domains among the ten domains of the Visual Domain Decathlon challenge. One can see that the subnetwork of these two highly related domains (one red, the other green) share a majority of building blocks from their overlapping paths, but there are still some differences.

Architecture blocks of two domains (ImageNet and Describable Textures) among the ten domains of the Visual Domain Decathlon challenge. Red and green path represents the subnetwork of ImageNet and Describable Textures, respectively. Dark pink nodes represent the blocks shared by multiple domains. Light pink nodes represent the blocks used by each path. The model is built based on MobileNet V3-like search space. The “dwb” block in the figure represents the dwbottleneck block. The “zero” block in the figure indicates the subnetwork skips that block.

Below we show the path similarity between domains among the ten domains of the Visual Domain Decathlon challenge. The similarity is measured by the Jaccard similarity score between the subnetworks of each domain, where higher means the paths are more similar. As one might expect, domains that are more similar share more nodes in the paths generated by MPNAS, which is also a signal of strong positive knowledge transfer. For example, the paths for similar domains (like ImageNet, CIFAR-100, and VGG Flower, which all include objects in natural images) have high scores, while the paths for dissimilar domains (like Daimler Pedestrian Classification and UCF101 Dynamic Images, which include pedestrians in grayscale images and human activity in natural color images, respectively) have low scores.

Confusion matrix for the Jaccard similarity score between the paths for the ten domains. Score value ranges from 0 to 1. A greater value indicates two paths share more nodes.

Training a Heterogeneous Multi-domain Model
In the second stage, the model resulting from MPNAS is trained from scratch for all domains. For this to work, it is necessary to define a unified objective function for all the domains. To successfully handle a large variety of domains, we designed an algorithm that adapts throughout the learning process such that losses are balanced across domains, called adaptive balanced domain prioritization (ABDP).

Below we show the accuracy, model size, and FLOPS of the model trained in different settings. We compare MPNAS to three other approaches:

• Domain independent NAS: Searching and training a model for each domain separately.
• Single path multi-head: Using a pre-trained model as a shared backbone for all domains with separated classification heads for each domain.
• Multi-head NAS: Searching a unified backbone architecture for all domains with separated classification heads for each domain.

From the results, we can observe that domain independent NAS requires building a bundle of models for each domain, resulting in a large model size. Although single path multi-head and multi-head NAS can reduce the model size and FLOPS significantly, forcing the domains to share the same backbone introduces negative knowledge transfer, decreasing overall accuracy.

MPNAS can build a small and efficient model while still maintaining high overall accuracy. The average accuracy of MPNAS is even 1.9% higher than the domain independent NAS approach since the model enables positive knowledge transfer. The figure below compares per domain top-1 accuracy of these approaches.

Top-1 accuracy of each Visual Decathlon domain.

Our evaluation shows that top-1 accuracy is improved from 69.96% to 71.78% (delta: +1.81%) by using ABDP as part of the search and training stages.

Top-1 accuracy for each Visual Decathlon domain trained by MPNAS with and without ABDP.

Future Work
We find MPNAS is an efficient solution to build a heterogeneous network to address the data imbalance, domain diversity, negative transfer, domain scalability, and large search space of possible parameter sharing strategies in MDL. By using a MobileNet-like search space, the resulting model is also mobile friendly. We are continuing to extend MPNAS for multi-task learning for tasks that are not compatible with existing search algorithms and hope others might use MPNAS to build a unified multi-domain model.

Acknowledgements
This work is made possible through a collaboration spanning several teams across Google. We’d like to acknowledge contributions from Junjie Ke, Joshua Greaves, Grace Chu, Ramin Mehran, Gabriel Bender, Xuhui Jia, Brendan Jou, Yukun Zhu, Luciano Sbaiz, Alec Go, Andrew Howard, Jeff Gilbert, Peyman Milanfar, and Ming-Tsuan Yang.

Read More

Posted by Arun Kandoor, Software Engineer, Google Research

The increasing demand for machine learning (ML) model inference on-device (for mobile devices, tablets, etc.) is driven by the rise of compute-intensive applications, the need to keep certain data on device for privacy and security reasons, and the desire to provide services when a network connection may not be available. However, on-device inference introduces a myriad of challenges, ranging from modeling to platform support requirements. These challenges relate to how different architectures are designed to optimize memory and computation, while still trying to maintain the quality of the model. From a platform perspective, the issue is identifying operations and building on top of them in a way that can generalize well across different product use cases.

In previous research, we combined a novel technique for generating embeddings (called projection-based embeddings) with efficient architectures like QRNN (pQRNN) and proved them to be competent for a number of classification problems. Augmenting these with distillation techniques provides an additional bump in end-to-end quality. Although this is an effective approach, it is not scalable to bigger and more extensive vocabularies (i.e., all possible Unicode or word tokens that can be fed to the model). Additionally, the output from the projection operation itself doesn’t contain trainable weights to take advantage of pre-training the model.

Token-free models presented in ByT5 are a good starting point for on-device modeling that can address pre-training and scalability issues without the need to increase the size of the model. This is possible because these approaches treat text inputs as a stream of bytes (each byte has a value that ranges from 0 to 255) that can reduce the vocabulary size for the embedding tables from ~30,000 to 256. Although ByT5 presents a compelling alternative for on-device modeling, going from word-level representation to byte stream representation increases the sequence lengths linearly; with an average word length of four characters and a single character having up to four bytes, the byte sequence length increases proportionally to the word length. This can lead to a significant increase in inference latency and computational costs.

We address this problem by developing and releasing three novel byte-stream sequence models for the SeqFlowLite library (ByteQRNN, ByteTransformer and ByteFunnelTransformer), all of which can be pre-trained on unsupervised data and can be fine-tuned for specific tasks. These models leverage recent innovations introduced by Charformer, including a fast character Transformer-based model that uses a gradient-based subword tokenization (GBST) approach to operate directly at the byte level, as well as a “soft” tokenization approach, which allows us to learn token boundaries and reduce sequence lengths. In this post, we focus on ByteQRNN and demonstrate that the performance of a pre-trained ByteQRNN model is comparable to BERT, despite being 300x smaller.

Sequence Model Architecture
We leverage pQRNN, ByT5 and Charformer along with platform optimizations, such as in-training quantization (which tracks minimum and maximum float values for model activations and weights for quantizing the inference model) that reduces model sizes by one-fourth, to develop an end-to-end model called ByteQRNN (shown below). First, we use a ByteSplitter operation to split the input string into a byte stream and feed it to a smaller embedding table that has a vocabulary size of 259 (256 + 3 additional meta tokens).

The output from the embedding layer is fed to the GBST layer, which is equipped with in-training quantization and combines byte-level representations with the efficiency of subword tokenization while enabling end-to-end learning of latent subwords. We “soft” tokenize the byte stream sequences by enumerating and combining each subword block length with scores (computed with a quantized dense layer) at each strided token position (i.e., at token positions that are selected at regular intervals). Next, we downsample the byte stream to manageable sequence length and feed it to the encoder layer.

The output from the GBST layer can be downsampled to a lower sequence length for efficient encoder computation or can be used by an encoder, like Funnel Transformer, which pools the query length and reduces the self-attention computation to create the ByteFunnelTransformer model. The encoder in the end-to-end model can be replaced with any other encoder layer, such as the Transformer from the SeqFlowLite library, to create a ByteTransformer model.

A diagram of a generic end-to-end sequence model using byte stream input. The ByteQRNN model uses a QRNN encoder from the SeqFlowLite library.

In addition to the input embeddings (i.e., the output from the embedding layer described above), we go a step further to build an effective sequence-to-sequence (seq2seq) model. We do so by taking ByteQRNN and adding a Transformer-based decoder model along with a quantized beam search (or tree exploration) to go with it. The quantized beam search module reduces the inference latency when generating decoder outputs by computing the most likely beams (i.e., possible output sequences) using the logarithmic sum of previous and current probabilities and returns the resulting top beams. Here the system uses a more efficient 8-bit integer (uint8) format, compared to a typical single-precision floating-point format (float32) model.

The decoder Transformer model uses a merged attention sublayer (MAtt) to reduce the complexity of the decoder self-attention from quadratic to linear, thereby lowering the end-to-end latency. For each decoding step, MAtt uses a fixed-size cache for decoder self-attention compared to the increasing cache size of a traditional transformer decoder. The following figure illustrates how the beam search module interacts with the decoder layer to generate output tokens on-device using an edge device (e.g., mobile phones, tablets, etc.).

A comparison of cloud server decoding and on-device (edge device) implementation. Left: Cloud server beam search employs a Transformer-based decoder model with quadratic time self-attention in float32, which has an increasing cache size for each decoding step. Right: The edge device implementation employs a quantized beam search module along with a fixed-size cache and a linear time self-attention computation.

Evaluation
After developing ByteQRNN, we evaluate its performance on the civil_comments dataset using the area under the curve (AUC) metric and compare it to a pre-trained ByteQRNN and BERT (shown below). We demonstrate that the fine-tuned ByteQRNN improves the overall quality and brings its performance closer to the BERT models, despite being 300x smaller. Since SeqFlowLite models support in-training quantization that reduces model sizes by one-fourth, the resulting models scale well to low-compute devices. We chose multilingual data sources that related to the task for pre-training both BERT and byte stream models to achieve the best possible performance.

Comparison of ByteQRNN with fine-tuned ByteQRNN and BERT on the civil_comments dataset.

Conclusion
Following up on our previous work with pQRNN, we evaluate byte stream models for on-device use to enable pre-training and thereby improve model performance for on-device deployment. We present an evaluation for ByteQRNN with and without pre-training and demonstrate that the performance of the pre-trained ByteQRNN is comparable to BERT, despite being 300x smaller. In addition to ByteQRNN, we are also releasing ByteTransformer and ByteFunnelTransformer, two models which use different encoders, along with the merged attention decoder model and the beam search driver to run the inference through the SeqFlowLite library. We hope these models will provide researchers and product developers with valuable resources for future on-device deployments.

Acknowledgements
We would like to thank Khoa Trinh, Jeongwoo Ko, Peter Young and Yicheng Fan for helping with open-sourcing and evaluating the model. Thanks to Prabhu Kaliamoorthi for all the brainstorming and ideation. Thanks to Vinh Tran, Jai Gupta and Yi Tay for their help with pre-training byte stream models. Thanks to Ruoxin Sang, Haoyu Zhang, Ce Zheng, Chuanhao Zhuge and Jieying Luo for helping with the TPU training. Many thanks to Erik Vee, Ravi Kumar and the Learn2Compress leadership for sponsoring the project and their support and encouragement. Finally, we would like to thank Tom Small for the animated figure used in this post.

Read More

Posted by Ehsan Amid, Research Scientist, and Rohan Anil, Principal Engineer, Google Research, Brain Team

While model design and training data are key ingredients in a deep neural network’s (DNN’s) success, less-often discussed is the specific optimization method used for updating the model parameters (weights). Training DNNs involves minimizing a loss function that measures the discrepancy between the ground truth labels and the model’s predictions. Training is carried out by backpropagation, which adjusts the model weights via gradient descent steps. Gradient descent, in turn, updates the weights by using the gradient (i.e., derivative) of the loss with respect to the weights.

The simplest weight update corresponds to stochastic gradient descent, which, in every step, moves the weights in the negative direction with respect to the gradients (with an appropriate step size, a.k.a. the learning rate). More advanced optimization methods modify the direction of the negative gradient before updating the weights by using information from the past steps and/or the local properties (such as the curvature information) of the loss function around the current weights. For instance, a momentum optimizer encourages moving along the average direction of past updates, and the AdaGrad optimizer scales each coordinate based on the past gradients. These optimizers are commonly known as first-order methods since they generally modify the update direction using only information from the first-order derivative (i.e., gradient). More importantly, the components of the weight parameters are treated independently from each other.

More advanced optimization, such as Shampoo and K-FAC, capture the correlations between gradients of parameters and have been shown to improve convergence, reducing the number of iterations and improving the quality of the solution. These methods capture information about the local changes of the derivatives of the loss, i.e., changes in gradients. Using this additional information, higher-order optimizers can discover much more efficient update directions for training models by taking into account the correlations between different groups of parameters. On the downside, calculating higher-order update directions is computationally more expensive than first-order updates. The operation uses more memory for storing statistics and involves matrix inversion, thus hindering the applicability of higher-order optimizers in practice.

In “LocoProp: Enhancing BackProp via Local Loss Optimization”, we introduce a new framework for training DNN models. Our new framework, LocoProp, conceives neural networks as a modular composition of layers. Generally, each layer in a neural network applies a linear transformation on its inputs, followed by a non-linear activation function. In the new construction, each layer is allotted its own weight regularizer, output target, and loss function. The loss function of each layer is designed to match the activation function of the layer. Using this formulation, training minimizes the local losses for a given mini-batch of examples, iteratively and in parallel across layers. Our method performs multiple local updates per batch of examples using a first-order optimizer (like RMSProp), which avoids computationally expensive operations such as the matrix inversions required for higher-order optimizers. However, we show that the combined local updates look rather like a higher-order update. Empirically, we show that LocoProp outperforms first-order methods on a deep autoencoder benchmark and performs comparably to higher-order optimizers, such as Shampoo and K-FAC, without the high memory and computation requirements.

Method
Neural networks are generally viewed as composite functions that transform model inputs into output representations, layer by layer. LocoProp adopts this view while decomposing the network into layers. In particular, instead of updating the weights of the layer to minimize the loss function at the output, LocoProp applies pre-defined local loss functions specific to each layer. For a given layer, the loss function is selected to match the activation function, e.g., a tanh loss would be selected for a layer with a tanh activation. Each layerwise loss measures the discrepancy between the layer’s output (for a given mini-batch of examples) and a notion of a target output for that layer. Additionally, a regularizer term ensures that the updated weights do not drift too far from the current values. The combined layerwise loss function (with a local target) plus regularizer is used as the new objective function for each layer.

Similar to backpropagation, LocoProp applies a forward pass to compute the activations. In the backward pass, LocoProp sets per neuron “targets” for each layer. Finally, LocoProp splits model training into independent problems across layers where several local updates can be applied to each layer’s weights in parallel.

Perhaps the simplest loss function one can think of for a layer is the squared loss. While the squared loss is a valid choice of a loss function, LocoProp takes into account the possible non-linearity of the activation functions of the layers and applies layerwise losses tailored to the activation function of each layer. This enables the model to emphasize regions at the input that are more important for the model prediction while deemphasizing the regions that do not affect the output as much. Below we show examples of tailored losses for the tanh and ReLU activation functions.

Loss functions induced by the (left) tanh and (right) ReLU activation functions. Each loss is more sensitive to the regions affecting the output prediction. For instance, ReLU loss is zero as long as both the prediction (â) and the target (a) are negative. This is because the ReLU function applied to any negative number equals zero.

After forming the objective in each layer, LocoProp updates the layer weights by repeatedly applying gradient descent steps on its objective. The update typically uses a first-order optimizer (like RMSProp). However, we show that the overall behavior of the combined updates closely resembles higher-order updates (shown below). Thus, LocoProp provides training performance close to what higher-order optimizers achieve without the high memory or computation needed for higher-order methods, such as matrix inverse operations. We show that LocoProp is a flexible framework that allows the recovery of well-known algorithms and enables the construction of new algorithms via different choices of losses, targets, and regularizers. LocoProp’s layerwise view of neural networks also allows updating the weights in parallel across layers.

Experiments
In our paper, we describe experiments on the deep autoencoder model, which is a commonly used baseline for evaluating the performance of optimization algorithms. We perform extensive tuning on multiple commonly used first-order optimizers, including SGD, SGD with momentum, AdaGrad, RMSProp, and Adam, as well as the higher-order Shampoo and K-FAC optimizers, and compare the results with LocoProp. Our findings indicate that the LocoProp method performs significantly better than first-order optimizers and is comparable to those of higher-order, while being significantly faster when run on a single GPU.

Train loss vs. number of epochs (left) and wall-clock time, i.e., the real time that passes during training, (right) for RMSProp, Shampoo, K-FAC, and LocoProp on the deep autoencoder model.

Summary and Future Directions
We introduced a new framework, called LocoProp, for optimizing deep neural networks more efficiently. LocoProp decomposes neural networks into separate layers with their own regularizer, output target, and loss function and applies local updates in parallel to minimize the local objectives. While using first-order updates for the local optimization problems, the combined updates closely resemble higher-order update directions, both theoretically and empirically.

LocoProp provides flexibility to choose the layerwise regularizers, targets, and loss functions. Thus, it allows the development of new update rules based on these choices. Our code for LocoProp is available online on GitHub. We are currently working on scaling up ideas induced by LocoProp to much larger scale models; stay tuned!

Acknowledgments
We would like to thank our co-author, Manfred K. Warmuth, for his critical contributions and inspiring vision. We would like to thank Sameer Agarwal for discussions looking at this work from a composite functions perspective, Vineet Gupta for discussions and development of Shampoo, Zachary Nado on K-FAC, Tom Small for development of the animation used in this blogpost and finally, Yonghui Wu and Zoubin Ghahramani for providing us with a nurturing research environment in the Google Brain Team.

Read More

Posted by Tuan Anh Nguyen, Google Assistant and Sourish Chaudhuri, Google Research

In natural conversations, we don’t say people’s names every time we speak to each other. Instead, we rely on contextual signaling mechanisms to initiate conversations, and eye contact is often all it takes. Google Assistant, now available in more than 95 countries and over 29 languages, has primarily relied on a hotword mechanism (“Hey Google” or “OK Google”) to help more than 700 million people every month get things done across Assistant devices. As virtual assistants become an integral part of our everyday lives, we’re developing ways to initiate conversations more naturally.

At Google I/O 2022, we announced Look and Talk, a major development in our journey to create natural and intuitive ways to interact with Google Assistant-powered home devices. This is the first multimodal, on-device Assistant feature that simultaneously analyzes audio, video, and text to determine when you are speaking to your Nest Hub Max. Using eight machine learning models together, the algorithm can differentiate intentional interactions from passing glances in order to accurately identify a user’s intent to engage with Assistant. Once within 5ft of the device, the user may simply look at the screen and talk to start interacting with the Assistant.

We developed Look and Talk in alignment with our AI Principles. It meets our strict audio and video processing requirements, and like our other camera sensing features, video never leaves the device. You can always stop, review and delete your Assistant activity at myactivity.google.com. These added layers of protection enable Look and Talk to work just for those who turn it on, while keeping your data safe.

Google Assistant relies on a number of signals to accurately determine when the user is speaking to it. On the right is a list of signals used with indicators showing when each signal is triggered based on the user’s proximity to the device and gaze direction.

Modeling Challenges
The journey of this feature began as a technical prototype built on top of models developed for academic research. Deployment at scale, however, required solving real-world challenges unique to this feature. It had to:

1. Support a range of demographic characteristics (e.g., age, skin tones).
2. Adapt to the ambient diversity of the real world, including challenging lighting (e.g., backlighting, shadow patterns) and acoustic conditions (e.g., reverberation, background noise).
3. Deal with unusual camera perspectives, since smart displays are commonly used as countertop devices and look up at the user(s), unlike the frontal faces typically used in research datasets to train models.
4. Run in real-time to ensure timely responses while processing video on-device.

The evolution of the algorithm involved experiments with approaches ranging from domain adaptation and personalization to domain-specific dataset development, field-testing and feedback, and repeated tuning of the overall algorithm.

Technology Overview
A Look and Talk interaction has three phases. In the first phase, Assistant uses visual signals to detect when a user is demonstrating an intent to engage with it and then “wakes up” to listen to their utterance. The second phase is designed to further validate and understand the user’s intent using visual and acoustic signals. If any signal in the first or second processing phases indicates that it isn’t an Assistant query, Assistant returns to standby mode. These two phases are the core Look and Talk functionality, and are discussed below. The third phase of query fulfillment is typical query flow, and is beyond the scope of this blog.

Phase One: Engaging with Assistant
The first phase of Look and Talk is designed to assess whether an enrolled user is intentionally engaging with Assistant. Look and Talk uses face detection to identify the user’s presence, filters for proximity using the detected face box size to infer distance, and then uses the existing Face Match system to determine whether they are enrolled Look and Talk users.

For an enrolled user within range, an custom eye gaze model determines whether they are looking at the device. This model estimates both the gaze angle and a binary gaze-on-camera confidence from image frames using a multi-tower convolutional neural network architecture, with one tower processing the whole face and another processing patches around the eyes. Since the device screen covers a region underneath the camera that would be natural for a user to look at, we map the gaze angle and binary gaze-on-camera prediction to the device screen area. To ensure that the final prediction is resilient to spurious individual predictions and involuntary eye blinks and saccades, we apply a smoothing function to the individual frame-based predictions to remove spurious individual predictions.

Eye-gaze prediction and post-processing overview.

We enforce stricter attention requirements before informing users that the system is ready for interaction to minimize false triggers, e.g., when a passing user briefly glances at the device. Once the user looking at the device starts speaking, we relax the attention requirement, allowing the user to naturally shift their gaze.

The final signal necessary in this processing phase checks that the Face Matched user is the active speaker. This is provided by a multimodal active speaker detection model that takes as input both video of the user’s face and the audio containing speech, and predicts whether they are speaking. A number of augmentation techniques (including RandAugment, SpecAugment, and augmenting with AudioSet sounds) helps improve prediction quality for the in-home domain, boosting end-feature performance by over 10%.The final deployed model is a quantized, hardware-accelerated TFLite model, which uses five frames of context for the visual input and 0.5 seconds for the audio input.

Active speaker detection model overview: The two-tower audiovisual model provides the “speaking” probability prediction for the face. The visual network auxiliary prediction pushes the visual network to be as good as possible on its own, improving the final multimodal prediction.

Phase Two: Assistant Starts Listening
In phase two, the system starts listening to the content of the user’s query, still entirely on-device, to further assess whether the interaction is intended for Assistant using additional signals. First, Look and Talk uses Voice Match to further ensure that the speaker is enrolled and matches the earlier Face Match signal. Then, it runs a state-of-the-art automatic speech recognition model on-device to transcribe the utterance.

The next critical processing step is the intent understanding algorithm, which predicts whether the user’s utterance was intended to be an Assistant query. This has two parts: 1) a model that analyzes the non-lexical information in the audio (i.e., pitch, speed, hesitation sounds) to determine whether the utterance sounds like an Assistant query, and 2) a text analysis model that determines whether the transcript is an Assistant request. Together, these filter out queries not intended for Assistant. It also uses contextual visual signals to determine the likelihood that the interaction was intended for Assistant.

Overview of the semantic filtering approach to determine if a user utterance is a query intended for the Assistant.

Finally, when the intent understanding model determines that the user utterance was likely meant for Assistant, Look and Talk moves into the fulfillment phase where it communicates with the Assistant server to obtain a response to the user’s intent and query text.

Performance, Personalization and UX
Each model that supports Look and Talk was evaluated and improved in isolation and then tested in the end-to-end Look and Talk system. The huge variety of ambient conditions in which Look and Talk operates necessitates the introduction of personalization parameters for algorithm robustness. By using signals obtained during the user’s hotword-based interactions, the system personalizes parameters to individual users to deliver improvements over the generalized global model. This personalization also runs entirely on-device.

Without a predefined hotword as a proxy for user intent, latency was a significant concern for Look and Talk. Often, a strong enough interaction signal does not occur until well after the user has started speaking, which can add hundreds of milliseconds of latency, and existing models for intent understanding add to this since they require complete, not partial, queries. To bridge this gap, Look and Talk completely forgoes streaming audio to the server, with transcription and intent understanding being on-device. The intent understanding models can work off of partial utterances. This results in an end-to-end latency comparable with current hotword-based systems.

The UI experience is based on user research to provide well-balanced visual feedback with high learnability. This is illustrated in the figure below.

Left: The spatial interaction diagram of a user engaging with Look and Talk. Right: The User Interface (UI) experience.

We developed a diverse video dataset with over 3,000 participants to test the feature across demographic subgroups. Modeling improvements driven by diversity in our training data improved performance for all subgroups.

Conclusion
Look and Talk represents a significant step toward making user engagement with Google Assistant as natural as possible. While this is a key milestone in our journey, we hope this will be the first of many improvements to our interaction paradigms that will continue to reimagine the Google Assistant experience responsibly. Our goal is to make getting help feel natural and easy, ultimately saving time so users can focus on what matters most.

Acknowledgements
This work involved collaborative efforts from a multidisciplinary team of software engineers, researchers, UX, and cross-functional contributors. Key contributors from Google Assistant include Alexey Galata, Alice Chuang‎, Barbara Wang, Britanie Hall, Gabriel Leblanc, Gloria McGee, Hideaki Matsui, James Zanoni, Joanna (Qiong) Huang, Krunal Shah, Kavitha Kandappan, Pedro Silva, Tanya Sinha, Tuan Nguyen, Vishal Desai, Will Truong‎, Yixing Cai‎, Yunfan Ye; from Research including Hao Wu, Joseph Roth, Sagar Savla, Sourish Chaudhuri, Susanna Ricco. Thanks to Yuan Yuan and Caroline Pantofaru for their leadership, and everyone on the Nest, Assistant, and Research teams who provided invaluable input toward the development of Look and Talk.

Read More

Posted by Maxim Tabachnyk, Staff Software Engineer and Stoyan Nikolov, Senior Engineering Manager, Google Research

The increasing complexity of code poses a key challenge to productivity in software engineering. Code completion has been an essential tool that has helped mitigate this complexity in integrated development environments (IDEs). Conventionally, code completion suggestions are implemented with rule-based semantic engines (SEs), which typically have access to the full repository and understand its semantic structure. Recent research has demonstrated that large language models (e.g., Codex and PaLM) enable longer and more complex code suggestions, and as a result, useful products have emerged (e.g., Copilot). However, the question of how code completion powered by machine learning (ML) impacts developer productivity, beyond perceived productivity and accepted suggestions, remains open.

Today we describe how we combined ML and SE to develop a novel Transformer-based hybrid semantic ML code completion, now available to internal Google developers. We discuss how ML and SEs can be combined by (1) re-ranking SE single token suggestions using ML, (2) applying single and multi-line completions using ML and checking for correctness with the SE, or (3) using single and multi-line continuation by ML of single token semantic suggestions. We compare the hybrid semantic ML code completion of 10k+ Googlers (over three months across eight programming languages) to a control group and see a 6% reduction in coding iteration time (time between builds and tests) and a 7% reduction in context switches (i.e., leaving the IDE) when exposed to single-line ML completion. These results demonstrate that the combination of ML and SEs can improve developer productivity. Currently, 3% of new code (measured in characters) is now generated from accepting ML completion suggestions.

Transformers for Completion
A common approach to code completion is to train transformer models, which use a self-attention mechanism for language understanding, to enable code understanding and completion predictions. We treat code similar to language, represented with sub-word tokens and a SentencePiece vocabulary, and use encoder-decoder transformer models running on TPUs to make completion predictions. The input is the code that is surrounding the cursor (~1000-2000 tokens) and the output is a set of suggestions to complete the current or multiple lines. Sequences are generated with a beam search (or tree exploration) on the decoder.

During training on Google’s monorepo, we mask out the remainder of a line and some follow-up lines, to mimic code that is being actively developed. We train a single model on eight languages (C++, Java, Python, Go, Typescript, Proto, Kotlin, and Dart) and observe improved or equal performance across all languages, removing the need for dedicated models. Moreover, we find that a model size of ~0.5B parameters gives a good tradeoff for high prediction accuracy with low latency and resource cost. The model strongly benefits from the quality of the monorepo, which is enforced by guidelines and reviews. For multi-line suggestions, we iteratively apply the single-line model with learned thresholds for deciding whether to start predicting completions for the following line.

Encoder-decoder transformer models are used to predict the remainder of the line or lines of code.

Re-rank Single Token Suggestions with ML
While a user is typing in the IDE, code completions are interactively requested from the ML model and the SE simultaneously in the backend. The SE typically only predicts a single token. The ML models we use predict multiple tokens until the end of the line, but we only consider the first token to match predictions from the SE. We identify the top three ML suggestions that are also contained in the SE suggestions and boost their rank to the top. The re-ranked results are then shown as suggestions for the user in the IDE.

In practice, our SEs are running in the cloud, providing language services (e.g., semantic completion, diagnostics, etc.) with which developers are familiar, and so we collocated the SEs to run on the same locations as the TPUs performing ML inference. The SEs are based on an internal library that offers compiler-like features with low latencies. Due to the design setup, where requests are done in parallel and ML is typically faster to serve (~40 ms median), we do not add any latency to completions. We observe a significant quality improvement in real usage. For 28% of accepted completions, the rank of the completion is higher due to boosting, and in 0.4% of cases it is worse. Additionally, we find that users type >10% fewer characters before accepting a completion suggestion.

Check Single / Multi-line ML Completions for Semantic Correctness
At inference time, ML models are typically unaware of code outside of their input window, and code seen during training might miss recent additions needed for completions in actively changing repositories. This leads to a common drawback of ML-powered code completion whereby the model may suggest code that looks correct, but doesn’t compile. Based on internal user experience research, this issue can lead to the erosion of user trust over time while reducing productivity gains.

We use SEs to perform fast semantic correctness checks within a given latency budget (<100ms for end-to-end completion) and use cached abstract syntax trees to enable a “full” structural understanding. Typical semantic checks include reference resolution (i.e., does this object exist), method invocation checks (e.g., confirming the method was called with a correct number of parameters), and assignability checks (to confirm the type is as expected).

For example, for the coding language Go, ~8% of suggestions contain compilation errors before semantic checks. However, the application of semantic checks filtered out 80% of uncompilable suggestions. The acceptance rate for single-line completions improved by 1.9x over the first six weeks of incorporating the feature, presumably due to increased user trust. As a comparison, for languages where we did not add semantic checking, we only saw a 1.3x increase in acceptance.

Language servers with access to source code and the ML backend are collocated on the cloud. They both perform semantic checking of ML completion suggestions.

Results
With 10k+ Google-internal developers using the completion setup in their IDE, we measured a user acceptance rate of 25-34%. We determined that the transformer-based hybrid semantic ML code completion completes >3% of code, while reducing the coding iteration time for Googlers by 6% (at a 90% confidence level). The size of the shift corresponds to typical effects observed for transformational features (e.g., key framework) that typically affect only a subpopulation, whereas ML has the potential to generalize for most major languages and engineers.

Providing Long Completions while Exploring APIs
We also tightly integrated the semantic completion with full line completion. When the dropdown with semantic single token completions appears, we display inline the single-line completions returned from the ML model. The latter represent a continuation of the item that is the focus of the dropdown. For example, if a user looks at possible methods of an API, the inline full line completions show the full method invocation also containing all parameters of the invocation.

Integrated full line completions by ML continuing the semantic dropdown completion that is in focus.

Suggestions of multiple line completions by ML.

Conclusion and Future Work
We demonstrate how the combination of rule-based semantic engines and large language models can be used to significantly improve developer productivity with better code completion. As a next step, we want to utilize SEs further, by providing extra information to ML models at inference time. One example can be for long predictions to go back and forth between the ML and the SE, where the SE iteratively checks correctness and offers all possible continuations to the ML model. When adding new features powered by ML, we want to be mindful to go beyond just “smart” results, but ensure a positive impact on productivity.

Acknowledgements
This research is the outcome of a two-year collaboration between Google Core and Google Research, Brain Team. Special thanks to Marc Rasi, Yurun Shen, Vlad Pchelin, Charles Sutton, Varun Godbole, Jacob Austin, Danny Tarlow, Benjamin Lee, Satish Chandra, Ksenia Korovina, Stanislav Pyatykh, Cristopher Claeys, Petros Maniatis, Evgeny Gryaznov, Pavel Sychev, Chris Gorgolewski, Kristof Molnar, Alberto Elizondo, Ambar Murillo, Dominik Schulz, David Tattersall, Rishabh Singh, Manzil Zaheer, Ted Ying, Juanjo Carin, Alexander Froemmgen and Marcus Revaj for their contributions.

Read More

Posted by Winnie Xu, Student Researcher and Kuang-Huei Lee, Software Engineer, Google Research, Brain Team

Current deep reinforcement learning (RL) methods can train specialist artificial agents that excel at decision-making on various individual tasks in specific environments, such as Go or StarCraft. However, little progress has been made to extend these results to generalist agents that would not only be capable of performing many different tasks, but also upon a variety of environments with potentially distinct embodiments.

Looking across recent progress in the fields of natural language processing, vision, and generative models (such as PaLM, Imagen, and Flamingo), we see that breakthroughs in making general-purpose models are often achieved by scaling up Transformer-based models and training them on large and semantically diverse datasets. It is natural to wonder, can a similar strategy be used in building generalist agents for sequential decision making? Can such models also enable fast adaptation to new tasks, similar to PaLM and Flamingo?

As an initial step to answer these questions, in our recent paper “Multi-Game Decision Transformers” we explore how to build a generalist agent to play many video games simultaneously. Our model trains an agent that can play 41 Atari games simultaneously at close-to-human performance and that can also be quickly adapted to new games via fine-tuning. This approach significantly improves upon the few existing alternatives to learning multi-game agents, such as temporal difference (TD) learning or behavioral cloning (BC).

A Multi-Game Decision Transformer (MGDT) can play multiple games at desired level of competency from training on a range of trajectories spanning all levels of expertise.

Don’t Optimize for Return, Just Ask for Optimality
In reinforcement learning, reward refers to the incentive signals that are relevant to completing a task, and return refers to cumulative rewards in a course of interactions between an agent and its surrounding environment. Traditional deep reinforcement learning agents (DQN, SimPLe, Dreamer, etc) are trained to optimize decisions to achieve the optimal return. At every time step, an agent observes the environment (some also consider the interactions that happened in the past) and decides what action to take to help itself achieve a higher return magnitude in future interactions.

In this work, we use Decision Transformers as our backbone approach to training an RL agent. A Decision Transformer is a sequence model that predicts future actions by considering past interactions between an agent and the surrounding environment, and (most importantly) a desired return to be achieved in future interactions. Instead of learning a policy to achieve high return magnitude as in traditional reinforcement learning, Decision Transformers map diverse experiences, ranging from expert-level to beginner-level, to their corresponding return magnitude during training. The idea is that training an agent on a range of experiences (from beginner to expert level) exposes the model to a wider range of variations in gameplay, which in turn helps it extract useful rules of gameplay that allow it to succeed under any circumstance. So during inference, the Decision Transformer can achieve any return value in the range it has seen during training, including the optimal return.

But, how do you know if a return is both optimal and stably achievable in a given environment? Previous applications of Decision Transformers relied on customized definitions of the desired return for each individual task, which required manually defining a plausible and informative range of scalar values that are appropriately interpretable signals for each specific game — a task that is non-trivial and rather unscalable. To address this issue, we instead model a distribution of return magnitudes based on past interactions with the environment during training. At inference time, we simply add an optimality bias that increases the probability of generating actions that are associated with higher returns.

To more comprehensively capture spatial-temporal patterns of agent-environment interactions, we also modified the Decision Transformer architecture to consider image patches instead of a global image representation. Patches allow the model to focus on local dynamics, which helps model game specific information in further detail.

These pieces together give us the backbone of Multi-Game Decision Transformers:

Each observation image is divided into a set of M patches of pixels which are denoted O. Return R, action a, and reward r follows these image patches in each input casual sequence. A Decision Transformer is trained to predict the next input (except for the image patches) to establish causality.

Training a Multi-Game Decision Transformer to Play 41 Games at Once
We train one Decision Transformer agent on a large (~1B) and broad set of gameplay experiences from 41 Atari games. In our experiments, this agent, which we call the Multi-Game Decision Transformer (MGDT), clearly outperforms existing reinforcement learning and behavioral cloning methods — by almost 2 times — on learning to play 41 games simultaneously and performs near human-level competency (100% in the following figure corresponds to the level of human gameplay). These results hold when comparing across training methods in both settings where a policy must be learned from static datasets (offline) as well as those where new data can be gathered from interacting with the environment (online).

Each bar is a combined score across 41 games, where 100% indicates human-level performance. Each blue bar is from a model trained on 41 games simultaneously, whereas each gray bar is from 41 specialist agents. Multi-Game Decision Transformer achieves human-level performance, significantly better than other multi-game agents, even comparable to specialist agents.

This result indicates that Decision Transformers are well-suited for multi-task, multi-environment, and multi-embodiment agents.

A concurrent work, “A Generalist Agent”, shows a similar result, demonstrating that large transformer-based sequence models can memorize expert behaviors very well across many more environments. In addition, their work and our work have nicely complementary findings: They show it’s possible to train across a wide range of environments beyond Atari games, while we show it’s possible and useful to train across a wide range of experiences.

In addition to the performance shown above, empirically we found that MGDT trained on a wide variety of experience is better than MDGT trained only on expert-level demonstrations or simply cloning demonstration behaviors.

Scaling Up Multi-Game Model Size to Achieve Better Performance
Argurably, scale has become the main driving force in many recent machine learning breakthroughs, and it is usually achieved by increasing the number of parameters in a transformer-based model. Our observation on Multi-Game Decision Transformers is similar: the performance increases predictably with larger model size. In particular, its performance appears to have not yet hit a ceiling, and compared to other learning systems performance gains are more significant with increases in model size.

Performance of Multi-Game Decision Transformer (shown by the blue line) increases predictably with larger model size, whereas other models do not.

Pre-trained Multi-Game Decision Transformers Are Fast Learners
Another benefit of MGDTs is that they can learn how to play a new game from very few gameplay demonstrations (which don’t need to all be expert-level). In that sense, MGDTs can be considered pre-trained models capable of being fine-tuned rapidly on small new gameplay data. Compared with other popular pre-training methods, it clearly shows consistent advantages in obtaining higher scores.

Multi-Game Decision Transformer pre-training (DT pre-training, shown in light blue) demonstrates consistent advantages over other popular models in adaptation to new tasks.

Where Is the Agent Looking?
In addition to the quantitative evaluation, it’s insightful (and fun) to visualize the agent’s behavior. By probing the attention heads, we find that the MGDT model consistently places weight in its field of view to areas of the observed images that contain meaningful game entities. We visualize the model’s attention when predicting the next action for various games and find it consistently attends to entities such as the agent’s on screen avatar, agent’s free movement space, non-agent objects, and key environment features. For example, in an interactive setting, having an accurate world model requires knowing how and when to focus on known objects (e.g., currently present obstacles) as well as expecting and/or planning over future unknowns (e.g., negative space). This diverse allocation of attention to many key components of each environment ultimately improves performance.

Here we can see the amount of weight the model places on each key asset of the game scene. Brighter red indicates more emphasis on that patch of pixels.

The Future of Large-Scale Generalist Agents
This work is an important step in demonstrating the possibility of training general-purpose agents across many environments, embodiments, and behavior styles. We have shown the benefit of increased scale on performance and the potential with further scaling. These findings seem to point to a generalization narrative similar to other domains like vision and language — we look forward to exploring the great potential of scaling data and learning from diverse experiences.

We look forward to future research towards developing performant agents for multi-environment and multi-embodiment settings. Our code and model checkpoints can soon be accessed here.

Acknowledgements
We’d like to thank all remaining authors of the paper including Igor Mordatch, Ofir Nachum Menjiao Yang, Lisa Lee, Daniel Freeman, Sergio Guadarrama, Ian Fischer, Eric Jang, Henryk Michalewski.

Read More

Posted by Akib Uddin, Product Manager and Andrew Sellergren, Software Engineer, Google Health

Every year, nearly a billion chest X-ray (CXR) images are taken globally to aid in the detection and management of health conditions ranging from collapsed lungs to infectious diseases. Generally, CXRs are cheaper and more accessible than other forms of medical imaging. However, existing challenges continue to impede the optimal use of CXRs. For example, in some areas, trained radiologists that can accurately interpret CXR images are in short supply. In addition, interpretation variability between experts, workflow differences between institutions, and the presence of rare conditions familiar only to subspecialists all contribute to making high-quality CXR interpretation a challenge.

Recent research has leveraged machine learning (ML) to explore potential solutions for some of these challenges. There is significant interest and effort devoted to building deep learning models that detect abnormalities in CXRs and improve access, accuracy, and efficiency to identify diseases and conditions that affect the heart and lungs. However, building robust CXR models requires large labeled training datasets, which can be prohibitively expensive and time-consuming to create. In some cases, such as working with underrepresented populations or studying rare medical conditions, only limited data are available. Additionally, CXR images vary in quality across populations, geographies, and institutions, making it difficult to build robust models that perform well globally.

In “Simplified Transfer Learning for Chest Radiography Models Using Less Data”, published in the journal Radiology, we describe how Google Health utilizes advanced ML methods to generate pre-trained “CXR networks” that can convert CXR images to embeddings (i.e., information-rich numerical vectors) to enable the development of CXR models using less data and fewer computational resources. We demonstrate that even with less data and compute, this approach has enabled performance comparable to state-of-the-art deep learning models across various prediction tasks. We are also excited to announce the release of CXR Foundation, a tool that utilizes our CXR-specific network to enable developers to create custom embeddings for their CXR images. We believe this work will help accelerate the development of CXR models, aiding in disease detection and contributing to more equitable health access throughout the world.

Developing a Chest X-ray Network
A common approach to building medical ML models is to pre-train a model on a generic task using non-medical datasets and then refine the model on a target medical task. This process of transfer learning may improve the target task performance or at least speed up convergence by applying the understanding of natural images to medical images. However, transfer learning may still require large labeled medical datasets for the refinement step.

Expanding on this standard approach, our system supports modeling CXR-specific tasks through a three-step model training setup composed of (1) generic image pre-training similar to traditional transfer learning, (2) CXR-specific pre-training, and (3) task-specific training. The first and third steps are common in ML: first pre-training on a large dataset and labels that are not specific to the desired task, and then fine-tuning on the task of interest.

We built a CXR-specific image classifier that employs supervised contrastive learning (SupCon). SupCon pulls together representations of images that have the same label (e.g., abnormal) and pushes apart representations of images that have a different label (e.g., one normal image and one abnormal image). We pre-trained this model on de-identified CXR datasets of over 800,000 images generated in partnership with Northwestern Medicine and Apollo Hospitals in the US and India, respectively. We then leveraged noisy abnormality labels from natural language processing of radiology reports to build our “CXR-specific” network.

This network creates embeddings (i.e., information-rich numerical vectors that can be used to distinguish classes from each other) that can more easily train models for specific medical prediction tasks, such as image finding (e.g., airspace opacity), clinical condition (e.g., tuberculosis), or patient outcome (e.g., hospitalization). For example, the CXR network can generate embeddings for every image in a given CXR dataset. For these images, the generated embeddings and the labels for the desired target task (such as tuberculosis) are used as examples to train a small ML model.

Left: Training a CXR model for a given task generally requires a large number of labeled images and a significant amount of computational resources to create a foundation of neural network layers. Right: With the CXR network and tool providing this foundation, each new task requires only a fraction of the labeled images, computational resources, and neural network parameters compared to rebuilding the entire network from scratch.

Effects of CXR Pre-training
We visualized these embedding layers at each step of the process using airspace opacity as an example (see the figure below). Before SupCon-based pre-training, there was poor separation of normal and abnormal CXR embeddings. After SupCon-based pre-training, the positive examples were grouped more closely together, and the negative examples more closely together as well, indicating that the model had identified that images from each category resembled themselves.

Visualizations of the t-distributed stochastic neighbor embedding for generic vs. CXR-specific network embeddings. Embeddings are information-rich numerical vectors that alone can distinguish classes from each other, in this case, airspace opacity positive vs. negative.

Our research suggests that adding the second stage of pre-training enables high-quality models to be trained with up to 600-fold less data in comparison to traditional transfer learning approaches that leverage pre-trained models on generic, non-medical datasets. We found this to be true regardless of model architecture (e.g., ResNet or EfficientNet) or dataset used for natural image pre-training (e.g., ImageNet or JFT-300M). With this approach, researchers and developers can significantly reduce dataset size requirements.

Top: In a deep learning model, the neural network contains multiple layers of artificial neurons, with the first layer taking the CXR image as input, intermediate layers doing additional computation, and the final layer making the classification (e.g., airspace opacity: present vs. absent). The embedding layer is usually one of the last layers. Bottom left: The traditional transfer learning approach involves a two-step training setup where a generic pre-trained network is optimized directly on a prediction task of interest. Our proposed three-step training setup generates a CXR network using a SupCon ML technique (step 2) before optimization for prediction tasks of interest (step 3). Bottom right: Using the embeddings involves either training smaller models (the first two strategies) or fine-tuning the whole network if there are sufficient data (strategy 3).

Results
After training the initial model, we measured performance using the area under the curve (AUC) metric with both linear and non-linear models applied to CXR embeddings; and a non-linear model produced by fine-tuning the entire network. On public datasets, such as ChestX-ray14 and CheXpert, our work substantially and consistently improved the data-accuracy tradeoff for models developed across a range of training dataset sizes and several findings. For example, when evaluating the tool’s ability to develop tuberculosis models, data efficiency gains were more striking: models trained on the embeddings of just 45 images achieved non-inferiority to radiologists in detecting tuberculosis on an external validation dataset. For both tuberculosis and severe COVID-19 outcomes, we show that non-linear classifiers trained on frozen embeddings outperformed a model that was fine-tuned on the entire dataset.

Comparing CXR-specific networks for transfer learning (red), with a baseline transfer learning approach (blue) across a variety of CXR abnormalities (top left), tuberculosis (bottom left), and COVID-19 outcomes (bottom right). This approach improves performance at the same dataset size, or reduces the dataset size required to reach the same performance. Interestingly, using the CXR network with simpler ML models that are faster to train (red) performs better than training the full network (black) at dataset sizes up to 85 images.

Conclusion and Future Work
To accelerate CXR modeling efforts with low data and computational requirements, we are releasing our CXR Foundation tool, along with scripts to train linear and nonlinear classifiers. Via these embeddings, this tool will allow researchers to jump-start CXR modeling efforts using simpler transfer learning methods. This approach can be particularly useful for predictive modeling using small datasets, and for adapting CXR models when there are distribution shifts in patient populations (whether over time or across different institutions). We are excited to continue working with partners, such as Northwestern Medicine and Apollo Hospitals, to explore the impact of this technology further. By enabling researchers with limited data and compute to develop CXR models, we’re hoping more developers can solve the most impactful problems for their populations.

Acknowledgements
Key contributors to this project at Google include Christina Chen, Yun Liu, Dilip Krishnan, Zaid Nabulsi, Atilla Kiraly, Arnav Agharwal, Eric Wu, Yuanzhen Li, Aaron Maschinot, Aaron Sarna, Jenny Huang, Marilyn Zhang, Charles Lau, Neeral Beladia, Daniel Tse, Krish Eswaran, and Shravya Shetty. Significant contributions and input were also made by collaborators Sreenivasa Raju Kalidindi, Mozziyar Etemadi, Florencia Garcia-Vicente, and David Melnick. For the ChestX-ray14 dataset, we thank the NIH Clinical Center for making it publicly available. The authors would also like to acknowledge many members of the Google Health Radiology and labeling software teams. 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; Craig Mermel and Akinori Mitani for providing feedback on the manuscript; Nicole Linton and Lauren Winer for feedback on the blogpost; and Tom Small for the animation.

Read More