Category: Offsites

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

There has been great progress towards adapting large language models (LLMs) to accommodate multimodal inputs for tasks including image captioning, visual question answering (VQA), and open vocabulary recognition. Despite such achievements, current state-of-the-art visual language models (VLMs) perform inadequately on visual information seeking datasets, such as Infoseek and OK-VQA, where external knowledge is required to answer the questions.

Examples of visual information seeking queries where external knowledge is required to answer the question. Images are taken from the OK-VQA dataset.

In “AVIS: Autonomous Visual Information Seeking with Large Language Models”, we introduce a novel method that achieves state-of-the-art results on visual information seeking tasks. Our method integrates LLMs with three types of tools: (i) computer vision tools for extracting visual information from images, (ii) a web search tool for retrieving open world knowledge and facts, and (iii) an image search tool to glean relevant information from metadata associated with visually similar images. AVIS employs an LLM-powered planner to choose tools and queries at each step. It also uses an LLM-powered reasoner to analyze tool outputs and extract key information. A working memory component retains information throughout the process.

An example of AVIS’s generated workflow for answering a challenging visual information seeking question. The input image is taken from the Infoseek dataset.

Comparison to previous work

Recent studies (e.g., Chameleon, ViperGPT and MM-ReAct) explored adding tools to LLMs for multimodal inputs. These systems follow a two-stage process: planning (breaking down questions into structured programs or instructions) and execution (using tools to gather information). Despite success in basic tasks, this approach often falters in complex real-world scenarios.

There has also been a surge of interest in applying LLMs as autonomous agents (e.g., WebGPT and ReAct). These agents interact with their environment, adapt based on real-time feedback, and achieve goals. However, these methods do not restrict the tools that can be invoked at each stage, leading to an immense search space. Consequently, even the most advanced LLMs today can fall into infinite loops or propagate errors. AVIS tackles this via guided LLM use, influenced by human decisions from a user study.

Informing LLM decision making with a user study

Many of the visual questions in datasets such as Infoseek and OK-VQA pose a challenge even for humans, often requiring the assistance of various tools and APIs. An example question from the OK-VQA dataset is shown below. We conducted a user study to understand human decision-making when using external tools.

We conducted a user study to understand human decision-making when using external tools. Image is taken from the OK-VQA dataset.

The users were equipped with an identical set of tools as our method, including PALI, PaLM, and web search. They received input images, questions, detected object crops, and buttons linked to image search results. These buttons offered diverse information about the detected object crops, such as knowledge graph entities, similar image captions, related product titles, and identical image captions.

We record user actions and outputs and use it as a guide for our system in two key ways. First, we construct a transition graph (shown below) by analyzing the sequence of decisions made by users. This graph defines distinct states and restricts the available set of actions at each state. For example, at the start state, the system can take only one of these three actions: PALI caption, PALI VQA, or object detection. Second, we use the examples of human decision-making to guide our planner and reasoner with relevant contextual instances to enhance the performance and effectiveness of our system.

AVIS transition graph.

General framework

Our approach employs a dynamic decision-making strategy designed to respond to visual information-seeking queries. Our system has three primary components. First, we have a planner to determine the subsequent action, including the appropriate API call and the query it needs to process. Second, we have a working memory that retains information about the results obtained from API executions. Last, we have a reasoner, whose role is to process the outputs from the API calls. It determines whether the obtained information is sufficient to produce the final response, or if additional data retrieval is required.

The planner undertakes a series of steps each time a decision is required regarding which tool to employ and what query to send to it. Based on the present state, the planner provides a range of potential subsequent actions. The potential action space may be so large that it makes the search space intractable. To address this issue, the planner refers to the transition graph to eliminate irrelevant actions. The planner also excludes the actions that have already been taken before and are stored in the working memory.

Next, the planner collects a set of relevant in-context examples that are assembled from the decisions previously made by humans during the user study. With these examples and the working memory that holds data collected from past tool interactions, the planner formulates a prompt. The prompt is then sent to the LLM, which returns a structured answer, determining the next tool to be activated and the query to be dispatched to it. This design allows the planner to be invoked multiple times throughout the process, thereby facilitating dynamic decision-making that gradually leads to answering the input query.

We employ a reasoner to analyze the output of the tool execution, extract the useful information and decide into which category the tool output falls: informative, uninformative, or final answer. Our method utilizes the LLM with appropriate prompting and in-context examples to perform the reasoning. If the reasoner concludes that it’s ready to provide an answer, it will output the final response, thus concluding the task. If it determines that the tool output is uninformative, it will revert back to the planner to select another action based on the current state. If it finds the tool output to be useful, it will modify the state and transfer control back to the planner to make a new decision at the new state.

AVIS employs a dynamic decision-making strategy to respond to visual information-seeking queries.

Results

We evaluate AVIS on Infoseek and OK-VQA datasets. As shown below, even robust visual-language models, such as OFA and PaLI, fail to yield high accuracy when fine-tuned on Infoseek. Our approach (AVIS), without fine-tuning, achieves 50.7% accuracy on the unseen entity split of this dataset.

AVIS visual question answering results on Infoseek dataset. AVIS achieves higher accuracy in comparison to previous baselines based on PaLI, PaLM and OFA.

Our results on the OK-VQA dataset are shown below. AVIS with few-shot in-context examples achieves an accuracy of 60.2%, higher than most of the previous works. AVIS achieves lower but comparable accuracy in comparison to the PALI model fine-tuned on OK-VQA. This difference, compared to Infoseek where AVIS outperforms fine-tuned PALI, is due to the fact that most question-answer examples in OK-VQA rely on common sense knowledge rather than on fine-grained knowledge. Therefore, PaLI is able to encode such generic knowledge in the model parameters and doesn’t require external knowledge.

Visual question answering results on A-OKVQA. AVIS achieves higher accuracy in comparison to previous works that use few-shot or zero-shot learning, including Flamingo, PaLI and ViperGPT. AVIS also achieves higher accuracy than most of the previous works that are fine-tuned on OK-VQA dataset, including REVEAL, ReVIVE, KAT and KRISP, and achieves results that are close to the fine-tuned PaLI model.

Conclusion

We present a novel approach that equips LLMs with the ability to use a variety of tools for answering knowledge-intensive visual questions. Our methodology, anchored in human decision-making data collected from a user study, employs a structured framework that uses an LLM-powered planner to dynamically decide on tool selection and query formation. An LLM-powered reasoner is tasked with processing and extracting key information from the output of the selected tool. Our method iteratively employs the planner and reasoner to leverage different tools until all necessary information required to answer the visual question is amassed.

Acknowledgements

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

Posted by Hussein Hazimeh, Research Scientist, Athena Team, and Riade Benbaki, Graduate Student at MIT

Modern neural networks have achieved impressive performance across a variety of applications, such as language, mathematical reasoning, and vision. However, these networks often use large architectures that require lots of computational resources. This can make it impractical to serve such models to users, especially in resource-constrained environments like wearables and smartphones. A widely used approach to mitigate the inference costs of pre-trained networks is to prune them by removing some of their weights, in a way that doesn’t significantly affect utility. In standard neural networks, each weight defines a connection between two neurons. So after weights are pruned, the input will propagate through a smaller set of connections and thus requires less computational resources.

Original network vs. a pruned network.

Pruning methods can be applied at different stages of the network’s training process: post, during, or before training (i.e., immediately after weight initialization). In this post, we focus on the post-training setting: given a pre-trained network, how can we determine which weights should be pruned? One popular method is magnitude pruning, which removes weights with the smallest magnitude. While efficient, this method doesn’t directly consider the effect of removing weights on the network’s performance. Another popular paradigm is optimization-based pruning, which removes weights based on how much their removal impacts the loss function. Although conceptually appealing, most existing optimization-based approaches seem to face a serious tradeoff between performance and computational requirements. Methods that make crude approximations (e.g., assuming a diagonal Hessian matrix) can scale well, but have relatively low performance. On the other hand, while methods that make fewer approximations tend to perform better, they appear to be much less scalable.

In “Fast as CHITA: Neural Network Pruning with Combinatorial Optimization”, presented at ICML 2023, we describe how we developed an optimization-based approach for pruning pre-trained neural networks at scale. CHITA (which stands for “Combinatorial Hessian-free Iterative Thresholding Algorithm”) outperforms existing pruning methods in terms of scalability and performance tradeoffs, and it does so by leveraging advances from several fields, including high-dimensional statistics, combinatorial optimization, and neural network pruning. For example, CHITA can be 20x to 1000x faster than state-of-the-art methods for pruning ResNet and improves accuracy by over 10% in many settings.

Overview of contributions

CHITA has two notable technical improvements over popular methods:

• Efficient use of second-order information: Pruning methods that use second-order information (i.e., relating to second derivatives) achieve the state of the art in many settings. In the literature, this information is typically used by computing the Hessian matrix or its inverse, an operation that is very difficult to scale because the Hessian size is quadratic with respect to the number of weights. Through careful reformulation, CHITA uses second-order information without having to compute or store the Hessian matrix explicitly, thus allowing for more scalability.
• Combinatorial optimization: Popular optimization-based methods use a simple optimization technique that prunes weights in isolation, i.e., when deciding to prune a certain weight they don’t take into account whether other weights have been pruned. This could lead to pruning important weights because weights deemed unimportant in isolation may become important when other weights are pruned. CHITA avoids this issue by using a more advanced, combinatorial optimization algorithm that takes into account how pruning one weight impacts others.

In the sections below, we discuss CHITA’s pruning formulation and algorithms.

A computation-friendly pruning formulation

There are many possible pruning candidates, which are obtained by retaining only a subset of the weights from the original network. Let k be a user-specified parameter that denotes the number of weights to retain. Pruning can be naturally formulated as a best-subset selection (BSS) problem: among all possible pruning candidates (i.e., subsets of weights) with only k weights retained, the candidate that has the smallest loss is selected.

Pruning as a BSS problem: among all possible pruning candidates with the same total number of weights, the best candidate is defined as the one with the least loss. This illustration shows four candidates, but this number is generally much larger.

Solving the pruning BSS problem on the original loss function is generally computationally intractable. Thus, similar to previous work, such as OBD and OBS, we approximate the loss with a quadratic function by using a second-order Taylor series, where the Hessian is estimated with the empirical Fisher information matrix. While gradients can be typically computed efficiently, computing and storing the Hessian matrix is prohibitively expensive due to its sheer size. In the literature, it is common to deal with this challenge by making restrictive assumptions on the Hessian (e.g., diagonal matrix) and also on the algorithm (e.g., pruning weights in isolation).

CHITA uses an efficient reformulation of the pruning problem (BSS using the quadratic loss) that avoids explicitly computing the Hessian matrix, while still using all the information from this matrix. This is made possible by exploiting the low-rank structure of the empirical Fisher information matrix. This reformulation can be viewed as a sparse linear regression problem, where each regression coefficient corresponds to a certain weight in the neural network. After obtaining a solution to this regression problem, coefficients set to zero will correspond to weights that should be pruned. Our regression data matrix is (n x p), where n is the batch (sub-sample) size and p is the number of weights in the original network. Typically n << p, so storing and operating with this data matrix is much more scalable than common pruning approaches that operate with the (p x p) Hessian.

CHITA reformulates the quadratic loss approximation, which requires an expensive Hessian matrix, as a linear regression (LR) problem. The LR’s data matrix is linear in p, which makes the reformulation more scalable than the original quadratic approximation.

Scalable optimization algorithms

CHITA reduces pruning to a linear regression problem under the following sparsity constraint: at most k regression coefficients can be nonzero. To obtain a solution to this problem, we consider a modification of the well-known iterative hard thresholding (IHT) algorithm. IHT performs gradient descent where after each update the following post-processing step is performed: all regression coefficients outside the Top-k (i.e., the k coefficients with the largest magnitude) are set to zero. IHT typically delivers a good solution to the problem, and it does so iteratively exploring different pruning candidates and jointly optimizing over the weights.

Due to the scale of the problem, standard IHT with constant learning rate can suffer from very slow convergence. For faster convergence, we developed a new line-search method that exploits the problem structure to find a suitable learning rate, i.e., one that leads to a sufficiently large decrease in the loss. We also employed several computational schemes to improve CHITA’s efficiency and the quality of the second-order approximation, leading to an improved version that we call CHITA++.

Experiments

We compare CHITA’s run time and accuracy with several state-of-the-art pruning methods using different architectures, including ResNet and MobileNet.

Run time: CHITA is much more scalable than comparable methods that perform joint optimization (as opposed to pruning weights in isolation). For example, CHITA’s speed-up can reach over 1000x when pruning ResNet.

Post-pruning accuracy: Below, we compare the performance of CHITA and CHITA++ with magnitude pruning (MP), Woodfisher (WF), and Combinatorial Brain Surgeon (CBS), for pruning 70% of the model weights. Overall, we see good improvements from CHITA and CHITA++.

Post-pruning accuracy of various methods on ResNet20. Results are reported for pruning 70% of the model weights.

Post-pruning accuracy of various methods on MobileNet. Results are reported for pruning 70% of the model weights.

Next, we report results for pruning a larger network: ResNet50 (on this network, some of the methods listed in the ResNet20 figure couldn’t scale). Here we compare with magnitude pruning and M-FAC. The figure below shows that CHITA achieves better test accuracy for a wide range of sparsity levels.

Test accuracy of pruned networks, obtained using different methods.

Conclusion, limitations, and future work

We presented CHITA, an optimization-based approach for pruning pre-trained neural networks. CHITA offers scalability and competitive performance by efficiently using second-order information and drawing on ideas from combinatorial optimization and high-dimensional statistics.

CHITA is designed for unstructured pruning in which any weight can be removed. In theory, unstructured pruning can significantly reduce computational requirements. However, realizing these reductions in practice requires special software (and possibly hardware) that support sparse computations. In contrast, structured pruning, which removes whole structures like neurons, may offer improvements that are easier to attain on general-purpose software and hardware. It would be interesting to extend CHITA to structured pruning.

Acknowledgements

This work is part of a research collaboration between Google and MIT. Thanks to Rahul Mazumder, Natalia Ponomareva, Wenyu Chen, Xiang Meng, Zhe Zhao, and Sergei Vassilvitskii for their help in preparing this post and the paper. Also thanks to John Guilyard for creating the graphics in this post.

Posted by Eltayeb Ahmed, Research Engineer, and Subhrajit Roy, Senior Research Scientist, Google Research

Reading has many benefits for young students, such as better linguistic and life skills, and reading for pleasure has been shown to correlate with academic success. Furthermore students have reported improved emotional wellbeing from reading, as well as better general knowledge and better understanding of other cultures. With the vast amount of reading material both online and off, finding age-appropriate, relevant and engaging content can be a challenging task, but helping students do so is a necessary step to engage them in reading. Effective recommendations that present students with relevant reading material helps keep students reading, and this is where machine learning (ML) can help.

ML has been widely used in building recommender systems for various types of digital content, ranging from videos to books to e-commerce items. Recommender systems are used across a range of digital platforms to help surface relevant and engaging content to users. In these systems, ML models are trained to suggest items to each user individually based on user preferences, user engagement, and the items under recommendation. These data provide a strong learning signal for models to be able to recommend items that are likely to be of interest, thereby improving user experience.

In “STUDY: Socially Aware Temporally Causal Decoder Recommender Systems”, we present a content recommender system for audiobooks in an educational setting taking into account the social nature of reading. We developed the STUDY algorithm in partnership with Learning Ally, an educational nonprofit, aimed at promoting reading in dyslexic students, that provides audiobooks to students through a school-wide subscription program. Leveraging the wide range of audiobooks in the Learning Ally library, our goal is to help students find the right content to help boost their reading experience and engagement. Motivated by the fact that what a person’s peers are currently reading has significant effects on what they would find interesting to read, we jointly process the reading engagement history of students who are in the same classroom. This allows our model to benefit from live information about what is currently trending within the student’s localized social group, in this case, their classroom.

Data

Learning Ally has a large digital library of curated audiobooks targeted at students, making it well-suited for building a social recommendation model to help improve student learning outcomes. We received two years of anonymized audiobook consumption data. All students, schools and groupings in the data were anonymized, only identified by a randomly generated ID not traceable back to real entities by Google. Furthermore all potentially identifiable metadata was only shared in an aggregated form, to protect students and institutions from being re-identified. The data consisted of time-stamped records of student’s interactions with audiobooks. For each interaction we have an anonymized student ID (which includes the student’s grade level and anonymized school ID), an audiobook identifier and a date. While many schools distribute students in a single grade across several classrooms, we leverage this metadata to make the simplifying assumption that all students in the same school and in the same grade level are in the same classroom. While this provides the foundation needed to build a better social recommender model, it’s important to note that this does not enable us to re-identify individuals, class groups or schools.

The STUDY algorithm

We framed the recommendation problem as a click-through rate prediction problem, where we model the conditional probability of a user interacting with each specific item conditioned on both 1) user and item characteristics and 2) the item interaction history sequence for the user at hand. Previous work suggests Transformer-based models, a widely used model class developed by Google Research, are well suited for modeling this problem. When each user is processed individually this becomes an autoregressive sequence modeling problem. We use this conceptual framework to model our data and then extend this framework to create the STUDY approach.

While this approach for click-through rate prediction can model dependencies between past and future item preferences for an individual user and can learn patterns of similarity across users at train time, it cannot model dependencies across different users at inference time. To recognise the social nature of reading and remediate this shortcoming we developed the STUDY model, which concatenates multiple sequences of books read by each student into a single sequence that collects data from multiple students in a single classroom.

However, this data representation requires careful diligence if it is to be modeled by transformers. In transformers, the attention mask is the matrix that controls which inputs can be used to inform the predictions of which outputs. The pattern of using all prior tokens in a sequence to inform the prediction of an output leads to the upper triangular attention matrix traditionally found in causal decoders. However, since the sequence fed into the STUDY model is not temporally ordered, even though each of its constituent subsequences is, a standard causal decoder is no longer a good fit for this sequence. When trying to predict each token, the model is not allowed to attend to every token that precedes it in the sequence; some of these tokens might have timestamps that are later and contain information that would not be available at deployment time.

In this figure we show the attention mask typically used in causal decoders. Each column represents an output and each column represents an output. A value of 1 (shown as blue) for a matrix entry at a particular position denotes that the model can observe the input of that row when predicting the output of the corresponding column, whereas a value of 0 (shown as white) denotes the opposite.

The STUDY model builds on causal transformers by replacing the triangular matrix attention mask with a flexible attention mask with values based on timestamps to allow attention across different subsequences. Compared to a regular transformer, which would not allow attention across different subsequences and would have a triangular matrix mask within sequence, STUDY maintains a causal triangular attention matrix within a sequence and has flexible values across sequences with values that depend on timestamps. Hence, predictions at any output point in the sequence are informed by all input points that occurred in the past relative to the current time point, regardless of whether they appear before or after the current input in the sequence. This causal constraint is important because if it is not enforced at train time, the model could potentially learn to make predictions using information from the future, which would not be available for a real world deployment.

In (a) we show a sequential autoregressive transformer with causal attention that processes each user individually; in (b) we show an equivalent joint forward pass that results in the same computation as (a); and finally, in (c) we show that by introducing new nonzero values (shown in purple) to the attention mask we allow information to flow across users. We do this by allowing a prediction to condition on all interactions with an earlier timestamp, irrespective of whether the interaction came from the same user or not.

<!–

In (a) we show a sequential autoregressive transformer with causal attention that processes each user individually; in (b) we show an equivalent joint forward pass that results in the same computation as (a); and finally, in (c) we show that by introducing new nonzero values (shown in purple) to the attention mask we allow information to flow across users. We do this by allowing a prediction to condition on all interactions with an earlier timestamp, irrespective of whether the interaction came from the same user or not.

–><!–

In (a) we show a sequential autoregressive transformer with causal attention that processes each user individually; in (b) we show an equivalent joint forward pass that results in the same computation as (a); and finally, in (c) we show that by introducing new nonzero values (shown in purple) to the attention mask we allow information to flow across users. We do this by allowing a prediction to condition on all interactions with an earlier timestamp, irrespective of whether the interaction came from the same user or not.

–>

Experiments

We used the Learning Ally dataset to train the STUDY model along with multiple baselines for comparison. We implemented an autoregressive click-through rate transformer decoder, which we refer to as “Individual”, a k-nearest neighbor baseline (KNN), and a comparable social baseline, social attention memory network (SAMN). We used the data from the first school year for training and we used the data from the second school year for validation and testing.

We evaluated these models by measuring the percentage of the time the next item the user actually interacted with was in the model’s top n recommendations, i.e., hits@n, for different values of n. In addition to evaluating the models on the entire test set we also report the models’ scores on two subsets of the test set that are more challenging than the whole data set. We observed that students will typically interact with an audiobook over multiple sessions, so simply recommending the last book read by the user would be a strong trivial recommendation. Hence, the first test subset, which we refer to as “non-continuation”, is where we only look at each model’s performance on recommendations when the students interact with books that are different from the previous interaction. We also observe that students revisit books they have read in the past, so strong performance on the test set can be achieved by restricting the recommendations made for each student to only the books they have read in the past. Although there might be value in recommending old favorites to students, much value from recommender systems comes from surfacing content that is new and unknown to the user. To measure this we evaluate the models on the subset of the test set where the students interact with a title for the first time. We name this evaluation subset “novel”.

We find that STUDY outperforms all other tested models across almost every single slice we evaluated against.

In this figure we compare the performance of four models, Study, Individual, KNN and SAMN. We measure the performance with hits@5, i.e., how likely the model is to suggest the next title the user read within the model’s top 5 recommendations. We evaluate the model on the entire test set (all) as well as the novel and non-continuation splits. We see STUDY consistently outperforms the other three models presented across all splits.

Importance of appropriate grouping

At the heart of the STUDY algorithm is organizing users into groups and doing joint inference over multiple users who are in the same group in a single forward pass of the model. We conducted an ablation study where we looked at the importance of the actual groupings used on the performance of the model. In our presented model we group together all students who are in the same grade level and school. We then experiment with groups defined by all students in the same grade level and district and also place all students in a single group with a random subset used for each forward pass. We also compare these models against the Individual model for reference.

We found that using groups that were more localized was more effective, with the school and grade level grouping outperforming the district and grade level grouping. This supports the hypothesis that the STUDY model is successful because of the social nature of activities such as reading — people’s reading choices are likely to correlate with the reading choices of those around them. Both of these models outperformed the other two models (single group and Individual) where grade level is not used to group students. This suggests that data from users with similar reading levels and interests is beneficial for performance.

Future work

This work is limited to modeling recommendations for user populations where the social connections are assumed to be homogenous. In the future it would be beneficial to model a user population where relationships are not homogeneous, i.e., where categorically different types of relationships exist or where the relative strength or influence of different relationships is known.

Acknowledgements

This work involved collaborative efforts from a multidisciplinary team of researchers, software engineers and educational subject matter experts. We thank our co-authors: Diana Mincu, Lauren Harrell, and Katherine Heller from Google. We also thank our colleagues at Learning Ally, Jeff Ho, Akshat Shah, Erin Walker, and Tyler Bastian, and our collaborators at Google, Marc Repnyek, Aki Estrella, Fernando Diaz, Scott Sanner, Emily Salkey and Lev Proleev.

Posted by Sandeep Tata, Software Engineer, Google Research, Athena Team

The last few years have seen rapid progress in systems that can automatically process complex business documents and turn them into structured objects. A system that can automatically extract data from documents, e.g., receipts, insurance quotes, and financial statements, has the potential to dramatically improve the efficiency of business workflows by avoiding error-prone, manual work. Recent models, based on the Transformer architecture, have shown impressive gains in accuracy. Larger models, such as PaLM 2, are also being leveraged to further streamline these business workflows. However, the datasets used in academic literature fail to capture the challenges seen in real-world use cases. Consequently, academic benchmarks report strong model accuracy, but these same models do poorly when used for complex real-world applications.

In “VRDU: A Benchmark for Visually-rich Document Understanding”, presented at KDD 2023, we announce the release of the new Visually Rich Document Understanding (VRDU) dataset that aims to bridge this gap and help researchers better track progress on document understanding tasks. We list five requirements for a good document understanding benchmark, based on the kinds of real-world documents for which document understanding models are frequently used. Then, we describe how most datasets currently used by the research community fail to meet one or more of these requirements, while VRDU meets all of them. We are excited to announce the public release of the VRDU dataset and evaluation code under a Creative Commons license.

Benchmark requirements

First, we compared state-of-the-art model accuracy (e.g., with FormNet and LayoutLMv2) on real-world use cases to academic benchmarks (e.g., FUNSD, CORD, SROIE). We observed that state-of-the-art models did not match academic benchmark results and delivered much lower accuracy in the real world. Next, we compared typical datasets for which document understanding models are frequently used with academic benchmarks and identified five dataset requirements that allow a dataset to better capture the complexity of real-world applications:

• Rich Schema: In practice, we see a wide variety of rich schemas for structured extraction. Entities have different data types (numeric, strings, dates, etc.) that may be required, optional, or repeated in a single document or may even be nested. Extraction tasks over simple flat schemas like (header, question, answer) do not reflect typical problems encountered in practice.
• Layout-Rich Documents: The documents should have complex layout elements. Challenges in practical settings come from the fact that documents may contain tables, key-value pairs, switch between single-column and double-column layout, have varying font-sizes for different sections, include pictures with captions and even footnotes. Contrast this with datasets where most documents are organized in sentences, paragraphs, and chapters with section headers — the kinds of documents that are typically the focus of classic natural language processing literature on long inputs.
• Diverse Templates: A benchmark should include different structural layouts or templates. It is trivial for a high-capacity model to extract from a particular template by memorizing the structure. However, in practice, one needs to be able to generalize to new templates/layouts, an ability that the train-test split in a benchmark should measure.
• High-Quality OCR: Documents should have high-quality Optical Character Recognition (OCR) results. Our aim with this benchmark is to focus on the VRDU task itself and to exclude the variability brought on by the choice of OCR engine.
• Token-Level Annotation: Documents should contain ground-truth annotations that can be mapped back to corresponding input text, so that each token can be annotated as part of the corresponding entity. This is in contrast with simply providing the text of the value to be extracted for the entity. This is key to generating clean training data where we do not have to worry about incidental matches to the given value. For instance, in some receipts, the ‘total-before-tax’ field may have the same value as the ‘total’ field if the tax amount is zero. Having token level annotations prevents us from generating training data where both instances of the matching value are marked as ground-truth for the ‘total’ field, thus producing noisy examples.

VRDU datasets and tasks

The VRDU dataset is a combination of two publicly available datasets, Registration Forms and Ad-Buy forms. These datasets provide examples that are representative of real-world use cases, and satisfy the five benchmark requirements described above.

The Ad-buy Forms dataset consists of 641 documents with political advertisement details. Each document is either an invoice or receipt signed by a TV station and a campaign group. The documents use tables, multi-columns, and key-value pairs to record the advertisement information, such as the product name, broadcast dates, total price, and release date and time.

The Registration Forms dataset consists of 1,915 documents with information about foreign agents registering with the US government. Each document records essential information about foreign agents involved in activities that require public disclosure. Contents include the name of the registrant, the address of related bureaus, the purpose of activities, and other details.

We gathered a random sample of documents from the public Federal Communications Commission (FCC) and Foreign Agents Registration Act (FARA) sites, and converted the images to text using Google Cloud’s OCR. We discarded a small number of documents that were several pages long and the processing did not complete in under two minutes. This also allowed us to avoid sending very long documents for manual annotation — a task that can take over an hour for a single document. Then, we defined the schema and corresponding labeling instructions for a team of annotators experienced with document-labeling tasks.

The annotators were also provided with a few sample labeled documents that we labeled ourselves. The task required annotators to examine each document, draw a bounding box around every occurrence of an entity from the schema for each document, and associate that bounding box with the target entity. After the first round of labeling, a pool of experts were assigned to review the results. The corrected results are included in the published VRDU dataset. Please see the paper for more details on the labeling protocol and the schema for each dataset.

Existing academic benchmarks (FUNSD, CORD, SROIE, Kleister-NDA, Kleister-Charity, DeepForm) fall-short on one or more of the five requirements we identified for a good document understanding benchmark. VRDU satisfies all of them. See our paper for background on each of these datasets and a discussion on how they fail to meet one or more of the requirements.

We built four different model training sets with 10, 50, 100, and 200 samples respectively. Then, we evaluated the VRDU datasets using three tasks (described below): (1) Single Template Learning, (2) Mixed Template Learning, and (3) Unseen Template Learning. For each of these tasks, we included 300 documents in the testing set. We evaluate models using the F1 score on the testing set.

• Single Template Learning (STL): This is the simplest scenario where the training, testing, and validation sets only contain a single template. This simple task is designed to evaluate a model’s ability to deal with a fixed template. Naturally, we expect very high F1 scores (0.90+) for this task.
• Mixed Template Learning (MTL): This task is similar to the task that most related papers use: the training, testing, and validation sets all contain documents belonging to the same set of templates. We randomly sample documents from the datasets and construct the splits to make sure the distribution of each template is not changed during sampling.
• Unseen Template Learning (UTL): This is the most challenging setting, where we evaluate if the model can generalize to unseen templates. For example, in the Registration Forms dataset, we train the model with two of the three templates and test the model with the remaining one. The documents in the training, testing, and validation sets are drawn from disjoint sets of templates. To our knowledge, previous benchmarks and datasets do not explicitly provide such a task designed to evaluate the model’s ability to generalize to templates not seen during training.

The objective is to be able to evaluate models on their data efficiency. In our paper, we compared two recent models using the STL, MTL, and UTL tasks and made three observations. First, unlike with other benchmarks, VRDU is challenging and shows that models have plenty of room for improvements. Second, we show that few-shot performance for even state-of-the-art models is surprisingly low with even the best models resulting in less than an F1 score of 0.60. Third, we show that models struggle to deal with structured repeated fields and perform particularly poorly on them.

Conclusion

We release the new Visually Rich Document Understanding (VRDU) dataset that helps researchers better track progress on document understanding tasks. We describe why VRDU better reflects practical challenges in this domain. We also present experiments showing that VRDU tasks are challenging, and recent models have substantial headroom for improvements compared to the datasets typically used in the literature with F1 scores of 0.90+ being typical. We hope the release of the VRDU dataset and evaluation code helps research teams advance the state of the art in document understanding.

Acknowledgements

Many thanks to Zilong Wang, Yichao Zhou, Wei Wei, and Chen-Yu Lee, who co-authored the paper along with Sandeep Tata. Thanks to Marc Najork, Riham Mansour and numerous partners across Google Research and the Cloud AI team for providing valuable insights. Thanks to John Guilyard for creating the animations in this post.

Posted by Fuzhao Xue, Research Intern, and Mostafa Dehghani, Research Scientist, Google

Adaptive computation refers to the ability of a machine learning system to adjust its behavior in response to changes in the environment. While conventional neural networks have a fixed function and computation capacity, i.e., they spend the same number of FLOPs for processing different inputs, a model with adaptive and dynamic computation modulates the computational budget it dedicates to processing each input, depending on the complexity of the input.

Adaptive computation in neural networks is appealing for two key reasons. First, the mechanism that introduces adaptivity provides an inductive bias that can play a key role in solving some challenging tasks. For instance, enabling different numbers of computational steps for different inputs can be crucial in solving arithmetic problems that require modeling hierarchies of different depths. Second, it gives practitioners the ability to tune the cost of inference through greater flexibility offered by dynamic computation, as these models can be adjusted to spend more FLOPs processing a new input.

Neural networks can be made adaptive by using different functions or computation budgets for various inputs. A deep neural network can be thought of as a function that outputs a result based on both the input and its parameters. To implement adaptive function types, a subset of parameters are selectively activated based on the input, a process referred to as conditional computation. Adaptivity based on the function type has been explored in studies on mixture-of-experts, where the sparsely activated parameters for each input sample are determined through routing.

Another area of research in adaptive computation involves dynamic computation budgets. Unlike in standard neural networks, such as T5, GPT-3, PaLM, and ViT, whose computation budget is fixed for different samples, recent research has demonstrated that adaptive computation budgets can improve performance on tasks where transformers fall short. Many of these works achieve adaptivity by using dynamic depth to allocate the computation budget. For example, the Adaptive Computation Time (ACT) algorithm was proposed to provide an adaptive computational budget for recurrent neural networks. The Universal Transformer extends the ACT algorithm to transformers by making the computation budget dependent on the number of transformer layers used for each input example or token. Recent studies, like PonderNet, follow a similar approach while improving the dynamic halting mechanisms.

In the paper “Adaptive Computation with Elastic Input Sequence”, we introduce a new model that utilizes adaptive computation, called AdaTape. This model is a Transformer-based architecture that uses a dynamic set of tokens to create elastic input sequences, providing a unique perspective on adaptivity in comparison to previous works. AdaTape uses an adaptive tape reading mechanism to determine a varying number of tape tokens that are added to each input based on input’s complexity. AdaTape is very simple to implement, provides an effective knob to increase the accuracy when needed, but is also much more efficient compared to other adaptive baselines because it directly injects adaptivity into the input sequence instead of the model depth. Finally, Adatape offers better performance on standard tasks, like image classification, as well as algorithmic tasks, while maintaining a favorable quality and cost tradeoff.

Adaptive computation transformer with elastic input sequence

AdaTape uses both the adaptive function types and a dynamic computation budget. Specifically, for a batch of input sequences after tokenization (e.g., a linear projection of non-overlapping patches from an image in the vision transformer), AdaTape uses a vector representing each input to dynamically select a variable-sized sequence of tape tokens.

AdaTape uses a bank of tokens, called a “tape bank”, to store all the candidate tape tokens that interact with the model through the adaptive tape reading mechanism. We explore two different methods for creating the tape bank: an input-driven bank and a learnable bank.

The general idea of the input-driven bank is to extract a bank of tokens from the input while employing a different approach than the original model tokenizer for mapping the raw input to a sequence of input tokens. This enables dynamic, on-demand access to information from the input that is obtained using a different point of view, e.g., a different image resolution or a different level of abstraction.

In some cases, tokenization in a different level of abstraction is not possible, thus an input-driven tape bank is not feasible, such as when it’s difficult to further split each node in a graph transformer. To address this issue, AdaTape offers a more general approach for generating the tape bank by using a set of trainable vectors as tape tokens. This approach is referred to as the learnable bank and can be viewed as an embedding layer where the model can dynamically retrieve tokens based on the complexity of the input example. The learnable bank enables AdaTape to generate a more flexible tape bank, providing it with the ability to dynamically adjust its computation budget based on the complexity of each input example, e.g., more complex examples retrieve more tokens from the bank, which let the model not only use the knowledge stored in the bank, but also spend more FLOPs processing it, since the input is now larger.

Finally, the selected tape tokens are appended to the original input and fed to the following transformer layers. For each transformer layer, the same multi-head attention is used across all input and tape tokens. However, two different feed-forward networks (FFN) are used: one for all tokens from the original input and the other for all tape tokens. We observed slightly better quality by using separate feed-forward networks for input and tape tokens.

An overview of AdaTape. For different samples, we pick a variable number of different tokens from the tape bank. The tape bank can be driven from input, e.g., by extracting some extra fine-grained information or it can be a set of trainable vectors. Adaptive tape reading is used to recursively select different sequences of tape tokens, with variable lengths, for different inputs. These tokens are then simply appended to inputs and fed to the transformer encoder.

AdaTape provides helpful inductive bias

We evaluate AdaTape on parity, a very challenging task for the standard Transformer, to study the effect of inductive biases in AdaTape. With the parity task, given a sequence 1s, 0s, and -1s, the model has to predict the evenness or oddness of the number of 1s in the sequence. Parity is the simplest non-counter-free or periodic regular language, but perhaps surprisingly, the task is unsolvable by the standard Transformer.

Evaluation on the parity task. The standard Transformer and Universal Transformer were unable to perform this task, both showing performance at the level of a random guessing baseline.

Despite being evaluated on short, simple sequences, both the standard Transformer and Universal Transformers were unable to perform the parity task as they are unable to maintain a counter within the model. However, AdaTape outperforms all baselines, as it incorporates a lightweight recurrence within its input selection mechanism, providing an inductive bias that enables the implicit maintenance of a counter, which is not possible in standard Transformers.

Evaluation on image classification

We also evaluate AdaTape on the image classification task. To do so, we trained AdaTape on ImageNet-1K from scratch. The figure below shows the accuracy of AdaTape and the baseline methods, including A-ViT, and the Universal Transformer ViT (UViT and U2T) versus their speed (measured as number of images, processed by each code, per second). In terms of quality and cost tradeoff, AdaTape performs much better than the alternative adaptive transformer baselines. In terms of efficiency, larger AdaTape models (in terms of parameter count) are faster than smaller baselines. Such results are consistent with the finding from previous work that shows that the adaptive model depth architectures are not well suited for many accelerators, like the TPU.

We evaluate AdaTape by training on ImageNet from scratch. For A-ViT, we not only report their results from the paper but also re-implement A-ViT by training from scratch, i.e., A-ViT(Ours).

A study of AdaTape’s behavior

In addition to its performance on the parity task and ImageNet-1K, we also evaluated the token selection behavior of AdaTape with an input-driven bank on the JFT-300M validation set. To better understand the model’s behavior, we visualized the token selection results on the input-driven bank as heatmaps, where lighter colors mean that position is more frequently selected. The heatmaps reveal that AdaTape more frequently picks the central patches. This aligns with our prior knowledge, as central patches are typically more informative — especially in the context of datasets with natural images, where the main object is in the middle of the image. This result highlights the intelligence of AdaTape, as it can effectively identify and prioritize more informative patches to improve its performance.

We visualize the tape token selection heatmap of AdaTape-B/32 (left) and AdaTape-B/16 (right). The hotter / lighter color means the patch at this position is more frequently selected.

Conclusion

AdaTape is characterized by elastic sequence lengths generated by the adaptive tape reading mechanism. This also introduces a new inductive bias that enables AdaTape to have the potential to solve tasks that are challenging for both standard transformers and existing adaptive transformers. By conducting comprehensive experiments on image recognition benchmarks, we demonstrate that AdaTape outperforms standard transformers and adaptive architecture transformers when computation is held constant.

Acknowledgments

One of the authors of this post, Mostafa Dehghani, is now at Google DeepMind.

Posted by Greg Corrado, Head of Health AI, Google Research, and Yossi Matias, VP, Engineering and Research, Google Research

Medicine is an inherently multimodal discipline. When providing care, clinicians routinely interpret data from a wide range of modalities including medical images, clinical notes, lab tests, electronic health records, genomics, and more. Over the last decade or so, AI systems have achieved expert-level performance on specific tasks within specific modalities — some AI systems processing CT scans, while others analyzing high magnification pathology slides, and still others hunting for rare genetic variations. The inputs to these systems tend to be complex data such as images, and they typically provide structured outputs, whether in the form of discrete grades or dense image segmentation masks. In parallel, the capacities and capabilities of large language models (LLMs) have become so advanced that they have demonstrated comprehension and expertise in medical knowledge by both interpreting and responding in plain language. But how do we bring these capabilities together to build medical AI systems that can leverage information from all these sources?

In today’s blog post, we outline a spectrum of approaches to bringing multimodal capabilities to LLMs and share some exciting results on the tractability of building multimodal medical LLMs, as described in three recent research papers. The papers, in turn, outline how to introduce de novo modalities to an LLM, how to graft a state-of-the-art medical imaging foundation model onto a conversational LLM, and first steps towards building a truly generalist multimodal medical AI system. If successfully matured, multimodal medical LLMs might serve as the basis of new assistive technologies spanning professional medicine, medical research, and consumer applications. As with our prior work, we emphasize the need for careful evaluation of these technologies in collaboration with the medical community and healthcare ecosystem.

A spectrum of approaches

Several methods for building multimodal LLMs have been proposed in recent months [1, 2, 3], and no doubt new methods will continue to emerge for some time. For the purpose of understanding the opportunities to bring new modalities to medical AI systems, we’ll consider three broadly defined approaches: tool use, model grafting, and generalist systems.

The spectrum of approaches to building multimodal LLMs range from having the LLM use existing tools or models, to leveraging domain-specific components with an adapter, to joint modeling of a multimodal model.

Tool use

In the tool use approach, one central medical LLM outsources analysis of data in various modalities to a set of software subsystems independently optimized for those tasks: the tools. The common mnemonic example of tool use is teaching an LLM to use a calculator rather than do arithmetic on its own. In the medical space, a medical LLM faced with a chest X-ray could forward that image to a radiology AI system and integrate that response. This could be accomplished via application programming interfaces (APIs) offered by subsystems, or more fancifully, two medical AI systems with different specializations engaging in a conversation.

This approach has some important benefits. It allows maximum flexibility and independence between subsystems, enabling health systems to mix and match products between tech providers based on validated performance characteristics of subsystems. Moreover, human-readable communication channels between subsystems maximize auditability and debuggability. That said, getting the communication right between independent subsystems can be tricky, narrowing the information transfer, or exposing a risk of miscommunication and information loss.

Model grafting

A more integrated approach would be to take a neural network specialized for each relevant domain, and adapt it to plug directly into the LLM — grafting the visual model onto the core reasoning agent. In contrast to tool use where the specific tool(s) used are determined by the LLM, in model grafting the researchers may choose to use, refine, or develop specific models during development. In two recent papers from Google Research, we show that this is in fact feasible. Neural LLMs typically process text by first mapping words into a vector embedding space. Both papers build on the idea of mapping data from a new modality into the input word embedding space already familiar to the LLM. The first paper, “Multimodal LLMs for health grounded in individual-specific data”, shows that asthma risk prediction in the UK Biobank can be improved if we first train a neural network classifier to interpret spirograms (a modality used to assess breathing ability) and then adapt the output of that network to serve as input into the LLM.

The second paper, “ELIXR: Towards a general purpose X-ray artificial intelligence system through alignment of large language models and radiology vision encoders”, takes this same tack, but applies it to full-scale image encoder models in radiology. Starting with a foundation model for understanding chest X-rays, already shown to be a good basis for building a variety of classifiers in this modality, this paper describes training a lightweight medical information adapter that re-expresses the top layer output of the foundation model as a series of tokens in the LLM’s input embeddings space. Despite fine-tuning neither the visual encoder nor the language model, the resulting system displays capabilities it wasn’t trained for, including semantic search and visual question answering.

Our approach to grafting a model works by training a medical information adapter that maps the output of an existing or refined image encoder into an LLM-understandable form.

Model grafting has a number of advantages. It uses relatively modest computational resources to train the adapter layers but allows the LLM to build on existing highly-optimized and validated models in each data domain. The modularization of the problem into encoder, adapter, and LLM components can also facilitate testing and debugging of individual software components when developing and deploying such a system. The corresponding disadvantages are that the communication between the specialist encoder and the LLM is no longer human readable (being a series of high dimensional vectors), and the grafting procedure requires building a new adapter for not just every domain-specific encoder, but also every revision of each of those encoders.

Generalist systems

The most radical approach to multimodal medical AI is to build one integrated, fully generalist system natively capable of absorbing information from all sources. In our third paper in this area, “Towards Generalist Biomedical AI”, rather than having separate encoders and adapters for each data modality, we build on PaLM-E, a recently published multimodal model that is itself a combination of a single LLM (PaLM) and a single vision encoder (ViT). In this set up, text and tabular data modalities are covered by the LLM text encoder, but now all other data are treated as an image and fed to the vision encoder.

Med-PaLM M is a large multimodal generative model that flexibly encodes and interprets biomedical data including clinical language, imaging, and genomics with the same model weights.

We specialize PaLM-E to the medical domain by fine-tuning the complete set of model parameters on medical datasets described in the paper. The resulting generalist medical AI system is a multimodal version of Med-PaLM that we call Med-PaLM M. The flexible multimodal sequence-to-sequence architecture allows us to interleave various types of multimodal biomedical information in a single interaction. To the best of our knowledge, it is the first demonstration of a single unified model that can interpret multimodal biomedical data and handle a diverse range of tasks using the same set of model weights across all tasks (detailed evaluations in the paper).

This generalist-system approach to multimodality is both the most ambitious and simultaneously most elegant of the approaches we describe. In principle, this direct approach maximizes flexibility and information transfer between modalities. With no APIs to maintain compatibility across and no proliferation of adapter layers, the generalist approach has arguably the simplest design. But that same elegance is also the source of some of its disadvantages. Computational costs are often higher, and with a unitary vision encoder serving a wide range of modalities, domain specialization or system debuggability could suffer.

The reality of multimodal medical AI

To make the most of AI in medicine, we’ll need to combine the strength of expert systems trained with predictive AI with the flexibility made possible through generative AI. Which approach (or combination of approaches) will be most useful in the field depends on a multitude of as-yet unassessed factors. Is the flexibility and simplicity of a generalist model more valuable than the modularity of model grafting or tool use? Which approach gives the highest quality results for a specific real-world use case? Is the preferred approach different for supporting medical research or medical education vs. augmenting medical practice? Answering these questions will require ongoing rigorous empirical research and continued direct collaboration with healthcare providers, medical institutions, government entities, and healthcare industry partners broadly. We look forward to finding the answers together.

Posted by Eleni Triantafillou, Research Scientist, and Malik Boudiaf, Student Researcher, Google

Deep learning has recently made tremendous progress in a wide range of problems and applications, but models often fail unpredictably when deployed in unseen domains or distributions. Source-free domain adaptation (SFDA) is an area of research that aims to design methods for adapting a pre-trained model (trained on a “source domain”) to a new “target domain”, using only unlabeled data from the latter.

Designing adaptation methods for deep models is an important area of research. While the increasing scale of models and training datasets has been a key ingredient to their success, a negative consequence of this trend is that training such models is increasingly computationally expensive, out of reach for certain practitioners and also harmful for the environment. One avenue to mitigate this issue is through designing techniques that can leverage and reuse already trained models for tackling new tasks or generalizing to new domains. Indeed, adapting models to new tasks is widely studied under the umbrella of transfer learning.

SFDA is a particularly practical area of this research because several real-world applications where adaptation is desired suffer from the unavailability of labeled examples from the target domain. In fact, SFDA is enjoying increasing attention [1, 2, 3, 4]. However, albeit motivated by ambitious goals, most SFDA research is grounded in a very narrow framework, considering simple distribution shifts in image classification tasks.

In a significant departure from that trend, we turn our attention to the field of bioacoustics, where naturally-occurring distribution shifts are ubiquitous, often characterized by insufficient target labeled data, and represent an obstacle for practitioners. Studying SFDA in this application can, therefore, not only inform the academic community about the generalizability of existing methods and identify open research directions, but can also directly benefit practitioners in the field and aid in addressing one of the biggest challenges of our century: biodiversity preservation.

In this post, we announce “In Search for a Generalizable Method for Source-Free Domain Adaptation”, appearing at ICML 2023. We show that state-of-the-art SFDA methods can underperform or even collapse when confronted with realistic distribution shifts in bioacoustics. Furthermore, existing methods perform differently relative to each other than observed in vision benchmarks, and surprisingly, sometimes perform worse than no adaptation at all. We also propose NOTELA, a new simple method that outperforms existing methods on these shifts while exhibiting strong performance on a range of vision datasets. Overall, we conclude that evaluating SFDA methods (only) on the commonly-used datasets and distribution shifts leaves us with a myopic view of their relative performance and generalizability. To live up to their promise, SFDA methods need to be tested on a wider range of distribution shifts, and we advocate for considering naturally-occurring ones that can benefit high-impact applications.

Distribution shifts in bioacoustics

Naturally-occurring distribution shifts are ubiquitous in bioacoustics. The largest labeled dataset for bird songs is Xeno-Canto (XC), a collection of user-contributed recordings of wild birds from across the world. Recordings in XC are “focal”: they target an individual captured in natural conditions, where the song of the identified bird is at the foreground. For continuous monitoring and tracking purposes, though, practitioners are often more interested in identifying birds in passive recordings (“soundscapes”), obtained through omnidirectional microphones. This is a well-documented problem that recent work shows is very challenging. Inspired by this realistic application, we study SFDA in bioacoustics using a bird species classifier that was pre-trained on XC as the source model, and several “soundscapes” coming from different geographical locations — Sierra Nevada (S. Nevada); Powdermill Nature Reserve, Pennsylvania, USA; Hawai’i; Caples Watershed, California, USA; Sapsucker Woods, New York, USA (SSW); and Colombia — as our target domains.

This shift from the focalized to the passive domain is substantial: the recordings in the latter often feature much lower signal-to-noise ratio, several birds vocalizing at once, and significant distractors and environmental noise, like rain or wind. In addition, different soundscapes originate from different geographical locations, inducing extreme label shifts since a very small portion of the species in XC will appear in a given location. Moreover, as is common in real-world data, both the source and target domains are significantly class imbalanced, because some species are significantly more common than others. In addition, we consider a multi-label classification problem since there may be several birds identified within each recording, a significant departure from the standard single-label image classification scenario where SFDA is typically studied.

Illustration of the “focal → soundscapes” shift. In the focalized domain, recordings are typically composed of a single bird vocalization in the foreground, captured with high signal-to-noise ratio (SNR), though there may be other birds vocalizing in the background. On the other hand, soundscapes contain recordings from omnidirectional microphones and can be composed of multiple birds vocalizing simultaneously, as well as environmental noises from insects, rain, cars, planes, etc.

Audio files		Focal domain		Soundscape domain1
Spectogram images

State-of-the-art SFDA models perform poorly on bioacoustics shifts

As a starting point, we benchmark six state-of-the-art SFDA methods on our bioacoustics benchmark, and compare them to the non-adapted baseline (the source model). Our findings are surprising: without exception, existing methods are unable to consistently outperform the source model on all target domains. In fact, they often underperform it significantly.

As an example, Tent, a recent method, aims to make models produce confident predictions for each example by reducing the uncertainty of the model’s output probabilities. While Tent performs well in various tasks, it doesn’t work effectively for our bioacoustics task. In the single-label scenario, minimizing entropy forces the model to choose a single class for each example confidently. However, in our multi-label scenario, there’s no such constraint that any class should be selected as being present. Combined with significant distribution shifts, this can cause the model to collapse, leading to zero probabilities for all classes. Other benchmarked methods like SHOT, AdaBN, Tent, NRC, DUST and Pseudo-Labelling, which are strong baselines for standard SFDA benchmarks, also struggle with this bioacoustics task.

Evolution of the test mean average precision (mAP), a standard metric for multilabel classification, throughout the adaptation procedure on the six soundscape datasets. We benchmark our proposed NOTELA and Dropout Student (see below), as well as SHOT, AdaBN, Tent, NRC, DUST and Pseudo-Labelling. Aside from NOTELA, all other methods fail to consistently improve the source model.

Introducing NOisy student TEacher with Laplacian Adjustment (NOTELA)

Nonetheless, a surprisingly positive result stands out: the less celebrated Noisy Student principle appears promising. This unsupervised approach encourages the model to reconstruct its own predictions on some target dataset, but under the application of random noise. While noise may be introduced through various channels, we strive for simplicity and use model dropout as the only noise source: we therefore refer to this approach as Dropout Student (DS). In a nutshell, it encourages the model to limit the influence of individual neurons (or filters) when making predictions on a specific target dataset.

DS, while effective, faces a model collapse issue on various target domains. We hypothesize this happens because the source model initially lacks confidence in those target domains. We propose improving DS stability by using the feature space directly as an auxiliary source of truth. NOTELA does this by encouraging similar pseudo-labels for nearby points in the feature space, inspired by NRC’s method and Laplacian regularization. This simple approach is visualized below, and consistently and significantly outperforms the source model in both audio and visual tasks.

NOTELA in action. The audio recordings are forwarded through the full model to obtain a first set of predictions, which are then refined through Laplacian regularization, a form of post-processing based on clustering nearby points. Finally, the refined predictions are used as targets for the noisy model to reconstruct.

Conclusion

The standard artificial image classification benchmarks have inadvertently limited our understanding of the true generalizability and robustness of SFDA methods. We advocate for broadening the scope and adopt a new assessment framework that incorporates naturally-occurring distribution shifts from bioacoustics. We also hope that NOTELA serves as a robust baseline to facilitate research in that direction. NOTELA’s strong performance perhaps points to two factors that can lead to developing more generalizable models: first, developing methods with an eye towards harder problems and second, favoring simple modeling principles. However, there is still future work to be done to pinpoint and comprehend existing methods’ failure modes on harder problems. We believe that our research represents a significant step in this direction, serving as a foundation for designing SFDA methods with greater generalizability.

Acknowledgements

One of the authors of this post, Eleni Triantafillou, is now at Google DeepMind. We are posting this blog post on behalf of the authors of the NOTELA paper: Malik Boudiaf, Tom Denton, Bart van Merriënboer, Vincent Dumoulin*, Eleni Triantafillou* (where * denotes equal contribution). We thank our co-authors for the hard work on this paper and the rest of the Perch team for their support and feedback.

¹Note that in this audio clip, the bird song is very faint; a common property in soundscape recordings where bird calls aren’t at the “foreground”. ^↩

Posted by Cat Armato, Program Manager, Google

Groups across Google actively pursue research in the field of machine learning (ML), ranging from theory and application. We build ML systems to solve deep scientific and engineering challenges in areas of language, music, visual processing, algorithm development, and more. We aim to build a more collaborative ecosystem with the broader ML research community through open-sourcing tools and datasets, publishing our work, and actively participating in conferences.

Google is proud to be a Diamond Sponsor of the 40th International Conference on Machine Learning (ICML 2023), a premier annual conference, which is being held this week in Honolulu, Hawaii. As a leader in ML research, Google has a strong presence at this year’s conference with over 120 accepted papers and active involvement in a number of workshops and tutorials. Google is also proud to be a Platinum Sponsor for both the LatinX in AI and Women in Machine Learning workshops. We look forward to sharing some of our extensive ML research and expanding our partnership with the broader ML research community.

Registered for ICML 2023? We hope you’ll visit the Google booth to learn more about the exciting work, creativity, and fun that goes into solving a portion of the field’s most interesting challenges. Visit the @GoogleAI Twitter account to find out about Google booth activities (e.g., demos and Q&A sessions). See Google DeepMind’s blog to learn about their technical participation at ICML 2023.

Take a look below to learn more about the Google research being presented at ICML 2023 (Google affiliations in bold).

Board and Organizing Committee

Google Research booth activities

Accepted papers

Tutorials

EXPO Day workshops

Google sponsored affinity workshops

Workshops

* Work done while at Google

Posted by Donald Martin, Jr., Technical Program Manager, Head of Societal Context Understanding Tools and Solutions (SCOUTS), Google Research

AI-related products and technologies are constructed and deployed in a societal context: that is, a dynamic and complex collection of social, cultural, historical, political and economic circumstances. Because societal contexts by nature are dynamic, complex, non-linear, contested, subjective, and highly qualitative, they are challenging to translate into the quantitative representations, methods, and practices that dominate standard machine learning (ML) approaches and responsible AI product development practices.

The first phase of AI product development is problem understanding, and this phase has tremendous influence over how problems (e.g., increasing cancer screening availability and accuracy) are formulated for ML systems to solve as well many other downstream decisions, such as dataset and ML architecture choice. When the societal context in which a product will operate is not articulated well enough to result in robust problem understanding, the resulting ML solutions can be fragile and even propagate unfair biases.

When AI product developers lack access to the knowledge and tools necessary to effectively understand and consider societal context during development, they tend to abstract it away. This abstraction leaves them with a shallow, quantitative understanding of the problems they seek to solve, while product users and society stakeholders — who are proximate to these problems and embedded in related societal contexts — tend to have a deep qualitative understanding of those same problems. This qualitative–quantitative divergence in ways of understanding complex problems that separates product users and society from developers is what we call the problem understanding chasm.

This chasm has repercussions in the real world: for example, it was the root cause of racial bias discovered by a widely used healthcare algorithm intended to solve the problem of choosing patients with the most complex healthcare needs for special programs. Incomplete understanding of the societal context in which the algorithm would operate led system designers to form incorrect and oversimplified causal theories about what the key problem factors were. Critical socio-structural factors, including lack of access to healthcare, lack of trust in the health care system, and underdiagnosis due to human bias, were left out while spending on healthcare was highlighted as a predictor of complex health need.

To bridge the problem understanding chasm responsibly, AI product developers need tools that put community-validated and structured knowledge of societal context about complex societal problems at their fingertips — starting with problem understanding, but also throughout the product development lifecycle. To that end, Societal Context Understanding Tools and Solutions (SCOUTS) — part of the Responsible AI and Human-Centered Technology (RAI-HCT) team within Google Research — is a dedicated research team focused on the mission to “empower people with the scalable, trustworthy societal context knowledge required to realize responsible, robust AI and solve the world’s most complex societal problems.” SCOUTS is motivated by the significant challenge of articulating societal context, and it conducts innovative foundational and applied research to produce structured societal context knowledge and to integrate it into all phases of the AI-related product development lifecycle. Last year we announced that Jigsaw, Google’s incubator for building technology that explores solutions to threats to open societies, leveraged our structured societal context knowledge approach during the data preparation and evaluation phases of model development to scale bias mitigation for their widely used Perspective API toxicity classifier. Going forward SCOUTS’ research agenda focuses on the problem understanding phase of AI-related product development with the goal of bridging the problem understanding chasm.

Bridging the AI problem understanding chasm

Bridging the AI problem understanding chasm requires two key ingredients: 1) a reference frame for organizing structured societal context knowledge and 2) participatory, non-extractive methods to elicit community expertise about complex problems and represent it as structured knowledge. SCOUTS has published innovative research in both areas.

A societal context reference frame

An essential ingredient for producing structured knowledge is a taxonomy for creating the structure to organize it. SCOUTS collaborated with other RAI-HCT teams (TasC, Impact Lab), Google DeepMind, and external system dynamics experts to develop a taxonomic reference frame for societal context. To contend with the complex, dynamic, and adaptive nature of societal context, we leverage complex adaptive systems (CAS) theory to propose a high-level taxonomic model for organizing societal context knowledge. The model pinpoints three key elements of societal context and the dynamic feedback loops that bind them together: agents, precepts, and artifacts.

• Agents: These can be individuals or institutions.
• Precepts: The preconceptions — including beliefs, values, stereotypes and biases — that constrain and drive the behavior of agents. An example of a basic precept is that “all basketball players are over 6 feet tall.” That limiting assumption can lead to failures in identifying basketball players of smaller stature.
• Artifacts: Agent behaviors produce many kinds of artifacts, including language, data, technologies, societal problems and products.

The relationships between these entities are dynamic and complex. Our work hypothesizes that precepts are the most critical element of societal context and we highlight the problems people perceive and the causal theories they hold about why those problems exist as particularly influential precepts that are core to understanding societal context. For example, in the case of racial bias in a medical algorithm described earlier, the causal theory precept held by designers was that complex health problems would cause healthcare expenditures to go up for all populations. That incorrect precept directly led to the choice of healthcare spending as the proxy variable for the model to predict complex healthcare need, which in turn led to the model being biased against Black patients who, due to societal factors such as lack of access to healthcare and underdiagnosis due to bias on average, do not always spend more on healthcare when they have complex healthcare needs. A key open question is how can we ethically and equitably elicit causal theories from the people and communities who are most proximate to problems of inequity and transform them into useful structured knowledge?

Illustrative version of societal context reference frame.

Taxonomic version of societal context reference frame.

Working with communities to foster the responsible application of AI to healthcare

Since its inception, SCOUTS has worked to build capacity in historically marginalized communities to articulate the broader societal context of the complex problems that matter to them using a practice called community based system dynamics (CBSD). System dynamics (SD) is a methodology for articulating causal theories about complex problems, both qualitatively as causal loop and stock and flow diagrams (CLDs and SFDs, respectively) and quantitatively as simulation models. The inherent support of visual qualitative tools, quantitative methods, and collaborative model building makes it an ideal ingredient for bridging the problem understanding chasm. CBSD is a community-based, participatory variant of SD specifically focused on building capacity within communities to collaboratively describe and model the problems they face as causal theories, directly without intermediaries. With CBSD we’ve witnessed community groups learn the basics and begin drawing CLDs within 2 hours.

Data 4 Black Lives community members learning system dynamics.

There is a huge potential for AI to improve medical diagnosis. But the safety, equity, and reliability of AI-related health diagnostic algorithms depends on diverse and balanced training datasets. An open challenge in the health diagnostic space is the dearth of training sample data from historically marginalized groups. SCOUTS collaborated with the Data 4 Black Lives community and CBSD experts to produce qualitative and quantitative causal theories for the data gap problem. The theories include critical factors that make up the broader societal context surrounding health diagnostics, including cultural memory of death and trust in medical care.

The figure below depicts the causal theory generated during the collaboration described above as a CLD. It hypothesizes that trust in medical care influences all parts of this complex system and is the key lever for increasing screening, which in turn generates data to overcome the data diversity gap.

Causal loop diagram of the health diagnostics data gap

These community-sourced causal theories are a first step to bridge the problem understanding chasm with trustworthy societal context knowledge.

Conclusion

As discussed in this blog, the problem understanding chasm is a critical open challenge in responsible AI. SCOUTS conducts exploratory and applied research in collaboration with other teams within Google Research, external community, and academic partners across multiple disciplines to make meaningful progress solving it. Going forward our work will focus on three key elements, guided by our AI Principles:

1. Increase awareness and understanding of the problem understanding chasm and its implications through talks, publications, and training.
2. Conduct foundational and applied research for representing and integrating societal context knowledge into AI product development tools and workflows, from conception to monitoring, evaluation and adaptation.
3. Apply community-based causal modeling methods to the AI health equity domain to realize impact and build society’s and Google’s capability to produce and leverage global-scale societal context knowledge to realize responsible AI.

SCOUTS flywheel for bridging the problem understanding chasm.

Acknowledgments

Thank you to John Guilyard for graphics development, everyone in SCOUTS, and all of our collaborators and sponsors.