Categories
Offsites

Talk like a graph: Encoding graphs for large language models

Imagine all the things around you — your friends, tools in your kitchen, or even the parts of your bike. They are all connected in different ways. In computer science, the term graph is used to describe connections between objects. Graphs consist of nodes (the objects themselves) and edges (connections between two nodes, indicating a relationship between them). Graphs are everywhere now. The internet itself is a giant graph of websites linked together. Even the knowledge search engines use is organized in a graph-like way.

Furthermore, consider the remarkable advancements in artificial intelligence — such as chatbots that can write stories in seconds, and even software that can interpret medical reports. This exciting progress is largely thanks to large language models (LLMs). New LLM technology is constantly being developed for different uses.

Since graphs are everywhere and LLM technology is on the rise, in “Talk like a Graph: Encoding Graphs for Large Language Models”, presented at ICLR 2024, we present a way to teach powerful LLMs how to better reason with graph information. Graphs are a useful way to organize information, but LLMs are mostly trained on regular text. The objective is to test different techniques to see what works best and gain practical insights. Translating graphs into text that LLMs can understand is a remarkably complex task. The difficulty stems from the inherent complexity of graph structures with multiple nodes and the intricate web of edges that connect them. Our work studies how to take a graph and translate it into a format that an LLM can understand. We also design a benchmark called GraphQA to study different approaches on different graph reasoning problems and show how to phrase a graph-related problem in a way that enables the LLM to solve the graph problem. We show that LLM performance on graph reasoning tasks varies on three fundamental levels: 1) the graph encoding method, 2) the nature of the graph task itself, and 3) interestingly, the very structure of the graph considered. These findings give us clues on how to best represent graphs for LLMs. Picking the right method can make the LLM up to 60% better at graph tasks!

Pictured, the process of encoding a graph as text using two different approaches and feeding the text and a question about the graph to the LLM.

Graphs as text

To be able to systematically find out what is the best way to translate a graph to text, we first design a benchmark called GraphQA. Think of GraphQA as an exam designed to evaluate powerful LLMs on graph-specific problems. We want to see how well LLMs can understand and solve problems that involve graphs in different setups. To create a comprehensive and realistic exam for LLMs, we don’t just use one type of graph, we use a mix of graphs ensuring breadth in the number of connections. This is mainly because different graph types make solving such problems easier or harder. This way, GraphQA can help expose biases in how an LLM thinks about the graphs, and the whole exam gets closer to a realistic setup that LLMs might encounter in the real world.

Overview of our framework for reasoning with graphs using LLMs.

GraphQA focuses on simple tasks related to graphs, like checking if an edge exists, calculating the number of nodes or edges, finding nodes that are connected to a specific node, and checking for cycles in a graph. These tasks might seem basic, but they require understanding the relationships between nodes and edges. By covering different types of challenges, from identifying patterns to creating new connections, GraphQA helps models learn how to analyze graphs effectively. These basic tasks are crucial for more complex reasoning on graphs, like finding the shortest path between nodes, detecting communities, or identifying influential nodes. Additionally, GraphQA includes generating random graphs using various algorithms like Erdős-Rényi, scale-free networks, Barabasi-Albert model, and stochastic block model, as well as simpler graph structures like paths, complete graphs, and star graphs, providing a diverse set of data for training.

When working with graphs, we also need to find ways to ask graph-related questions that LLMs can understand. Prompting heuristics are different strategies for doing this. Let’s break down the common ones:

  • Zero-shot: simply describe the task (“Is there a cycle in this graph?”) and tell the LLM to go for it. No examples provided.
  • Few-shot: This is like giving the LLM a mini practice test before the real deal. We provide a few example graph questions and their correct answers.
  • Chain-of-Thought: Here, we show the LLM how to break down a problem step-by-step with examples. The goal is to teach it to generate its own “thought process” when faced with new graphs.
  • Zero-CoT: Similar to CoT, but instead of training examples, we give the LLM a simple prompt, like “Let’s think step-by-step,” to trigger its own problem-solving breakdown.
  • BAG (build a graph): This is specifically for graph tasks. We add the phrase “Let’s build a graph…” to the description, helping the LLM focus on the graph structure.

We explored different ways to translate graphs into text that LLMs can work with. Our key questions were:

  • Node encoding: How do we represent individual nodes? Options tested include simple integers, common names (people, characters), and letters.
  • Edge encoding: How do we describe the relationships between nodes? Methods involved parenthesis notation, phrases like “are friends”, and symbolic representations like arrows.

Various node and edge encodings were combined systematically. This led to functions like the ones in the following figure:

Examples of graph encoding functions used to encode graphs via text.

Analysis and results

We carried out three key experiments: one to test how LLMs handle graph tasks, and two to understand how the size of the LLM and different graph shapes affected performance. We run all our experiments on GraphQA.

How LLMs handle graph tasks

In this experiment, we tested how well pre-trained LLMs tackle graph problems like identifying connections, cycles, and node degrees. Here is what we learned:

  • LLMs struggle: On most of these basic tasks, LLMs did not do much better than a random guess.
  • Encoding matters significantly: How we represent the graph as text has a great effect on LLM performance. The “incident” encoding excelled for most of the tasks in general.

Our results are summarized in the following chart.

Comparison of various graph encoder functions based on their accuracy on different graph tasks. The main conclusion from this figure is that the graph encoding functions matter significantly.

Bigger is (usually) better

In this experiment, we wanted to see if the size of the LLM (in terms of the number of parameters) affects how well they can handle graph problems. For that, we tested the same graph tasks on the XXS, XS, S, and L sizes of PaLM 2. Here is a summary of our findings:

  • In general, bigger models did better on graph reasoning tasks. It seems like the extra parameters gave them space to learn more complex patterns.
  • Oddly, size didn’t matter as much for the “edge existence” task (finding out if two nodes in a graph are connected).
  • Even the biggest LLM couldn’t consistently beat a simple baseline solution on the cycle check problem (finding out if a graph contains a cycle or not). This shows LLMs still have room to improve with certain graph tasks.
Effect of model capacity on graph reasoning task for PaLM 2-XXS, XS, S, and L.

Do different graph shapes confuse LLMs

We wondered if the “shape” of a graph (how nodes are connected) influences how well LLMs can solve problems on it. Think of the following figure as different examples of graph shapes.

Samples of graphs generated with different graph generators from GraphQA. ER, BA, SBM, and SFN refers to Erdős–Rényi, Barabási–Albert, Stochastic Block Model, and Scale-Free Network respectively.

We found that graph structure has a big impact on LLM performance. For example, in a task asking if a cycle exists, LLMs did great on tightly interconnected graphs (cycles are common there) but struggled on path graphs (where cycles never happen). Interestingly, providing some mixed examples helped it adapt. For instance, for cycle check, we added some examples containing a cycle and some examples with no cycles as few-shot examples in our prompt. Similar patterns occurred with other tasks.

Comparing different graph generators on different graph tasks. The main observation here is that graph structure has a significant impact on the LLM’s performance. ER, BA, SBM, and SFN refers to Erdős–Rényi, Barabási–Albert, Stochastic Block Model, and Scale-Free Network respectively.

Conclusion

In short, we dug deep into how to best represent graphs as text so LLMs can understand them. We found three major factors that make a difference:

  • How to translate the graph to text: how we represent the graph as text significantly influences LLM performance. The incident encoding excelled for most of the tasks in general..
  • Task type: Certain types of graph questions tend to be harder for LLMs, even with a good translation from graph to text.
  • Graph structure: Surprisingly, the “shape” of the graph that on which we do inference (dense with connections, sparse, etc.) influences how well an LLM does.

This study revealed key insights about how to prepare graphs for LLMs. The right encoding techniques can significantly boost an LLM’s accuracy on graph problems (ranging from around 5% to over 60% improvement). Our new benchmark, GraphQA, will help drive further research in this area.

Acknowledgements

We would like to express our gratitude to our co-author, Jonathan Halcrow, for his valuable contributions to this work. We express our sincere gratitude to Anton Tsitsulin, Dustin Zelle, Silvio Lattanzi, Vahab Mirrokni, and the entire graph mining team at Google Research, for their insightful comments, thorough proofreading, and constructive feedback which greatly enhanced the quality of our work. We would also like to extend special thanks to Tom Small for creating the animation used in this post.

Categories
Offsites

Cappy: Outperforming and boosting large multi-task language models with a small scorer

Large language model (LLM) advancements have led to a new paradigm that unifies various natural language processing (NLP) tasks within an instruction-following framework. This paradigm is exemplified by recent multi-task LLMs, such as T0, FLAN, and OPT-IML. First, multi-task data is gathered with each task following a task-specific template, where each labeled example is converted into an instruction (e.g., Put the concepts together to form a sentence: ski, mountain, skier) paired with a corresponding response (e.g., Skier skis down the mountain). These instruction-response pairs are used to train the LLM, resulting in a conditional generation model that takes an instruction as input and generates a response. Moreover, multi-task LLMs have exhibited remarkable task-wise generalization capabilities as they can address unseen tasks by understanding and solving brand-new instructions.

The demonstration of the instruction-following pre-training of multi-task LLMs, e.g., FLAN. Pre-training tasks under this paradigm improves the performance for unseen tasks.

Due to the complexity of understanding and solving various tasks solely using instructions, the size of multi-task LLMs typically spans from several billion parameters to hundreds of billions (e.g., FLAN-11B, T0-11B and OPT-IML-175B). As a result, operating such sizable models poses significant challenges because they demand considerable computational power and impose substantial requirements on the memory capacities of GPUs and TPUs, making their training and inference expensive and inefficient. Extensive storage is required to maintain a unique LLM copy for each downstream task. Moreover, the most powerful multi-task LLMs (e.g., FLAN-PaLM-540B) are closed-sourced, making them impossible to be adapted. However, in practical applications, harnessing a single multi-task LLM to manage all conceivable tasks in a zero-shot manner remains difficult, particularly when dealing with complex tasks, personalized tasks and those that cannot be succinctly defined using instructions. On the other hand, the size of downstream training data is usually insufficient to train a model well without incorporating rich prior knowledge. Hence, it is long desired to adapt LLMs with downstream supervision while bypassing storage, memory, and access issues.

Certain parameter-efficient tuning strategies, including prompt tuning and adapters, substantially diminish storage requirements, but they still perform back-propagation through LLM parameters during the tuning process, thereby keeping their memory demands high. Additionally, some in-context learning techniques circumvent parameter tuning by integrating a limited number of supervised examples into the instruction. However, these techniques are constrained by the model’s maximum input length, which permits only a few samples to guide task resolution.

In “Cappy: Outperforming and Boosting Large Multi-Task LMs with a Small Scorer”, presented at NeurIPS 2023, we propose a novel approach that enhances the performance and efficiency of multi-task LLMs. We introduce a lightweight pre-trained scorer, Cappy, based on continual pre-training on top of RoBERTa with merely 360 million parameters. Cappy takes in an instruction and a candidate response as input, and produces a score between 0 and 1, indicating an estimated correctness of the response with respect to the instruction. Cappy functions either independently on classification tasks or serves as an auxiliary component for LLMs, boosting their performance. Moreover, Cappy efficiently enables downstream supervision without requiring any finetuning, which avoids the need for back-propagation through LLM parameters and reduces memory requirements. Finally, adaptation with Cappy doesn’t require access to LLM parameters as it is compatible with closed-source multi-task LLMs, such as those only accessible via WebAPIs.

Cappy takes an instruction and response pair as input and outputs a score ranging from 0 to 1, indicating an estimation of the correctness of the response with respect to the instruction.

Pre-training

We begin with the same dataset collection, which includes 39 diverse datasets from PromptSource that were used to train T0. This collection encompasses a wide range of task types, such as question answering, sentiment analysis, and summarization. Each dataset is associated with one or more templates that convert each instance from the original datasets into an instruction paired with its ground truth response.

Cappy’s regression modeling requires each pre-training data instance to include an instruction-response pair along with a correctness annotation for the response, so we produce a dataset with correctness annotations that range from 0 to 1. For every instance within a generation task, we leverage an existing multi-task LLM to generate multiple responses by sampling, conditioned on the given instruction. Subsequently, we assign an annotation to the pair formed by the instruction and every response, using the similarity between the response and the ground truth response of the instance. Specifically, we employ Rouge-L, a commonly-used metric for measuring overall multi-task performance that has demonstrated a strong alignment with human evaluation, to calculate this similarity as a form of weak supervision.

As a result, we obtain an effective regression dataset of 160 million instances paired with correctness score annotations. The final Cappy model is the result of continuous pre-training using the regression dataset on top of the RoBERTa model. The pre-training of Cappy is conducted on Google’s TPU-v4, with RedCoast, a lightweight toolkit for automating distributed training.

Data augmentation with a multi-task LLM to construct a weakly supervised regression dataset for Cappy’s pre-training and fine-tuning.

Applying Cappy

Cappy solves practical tasks within a candidate-selection mechanism. More specifically, given an instruction and a set of candidate responses, Cappy produces a score for each candidate response. This is achieved by inputting the instruction alongside each individual response, and then assigning the response with the highest score as its prediction. In classification tasks, all candidate responses are inherently predefined. For example, for an instruction of a sentiment classification task (e.g., “Based on this review, would the user recommend this product?: ‘Stunning even for the non-gamer.’”), the candidate responses are “Yes” or “No”. In such scenarios, Cappy functions independently. On the other hand, in generation tasks, candidate responses are not pre-defined, requiring an existing multi-task LLM to yield the candidate responses. In this case, Cappy serves as an auxiliary component of the multi-task LLM, enhancing its decoding.

Adapting multi-task LLMs with Cappy

When there is available downstream training data, Cappy enables effective and efficient adaptation of multi-task LLMs on downstream tasks. Specifically, we fine-tune Cappy to integrate downstream task information into LLM predictions. This process involves creating a separate regression dataset specific to the downstream training data with the same data annotation process used to construct the pre-training data. As a result, the fine-tuned Cappy collaborates with a multi-task LLM, boosting the LLM’s performance on the downstream task.

In contrast to other LLM tuning strategies, adapting LLMs with Cappy significantly reduces the high demand for device memory as it avoids the need for back-propagation through LLM parameters for downstream tasks. Moreover, Cappy adaptation does not rely on the access to LLM parameters, making it compatible with closed-source multi-task LLMs, such as the ones only accessible via WebAPIs. Compared with in-context learning approaches, which circumvent model tuning by attaching training examples to the instruction prefix, Cappy is not restricted by the LLM’s maximum input length. Thus, Cappy can incorporate an unlimited number of downstream training examples. Cappy can also be applied with other adaptation methods, such as fine-tuning and in-context learning, further boosting their overall performance.

Downstream adaptation comparison between Cappy and approaches that rely on an LLM’s parameters, such as fine-tuning and prompt tuning. Cappy’s application enhances multi-task LLMs.

Results

We assess Cappy’s performance across eleven held-out language understanding classification tasks from PromptSource. We demonstrate that Cappy, with 360M parameters, outperforms OPT-175B and OPT-IML-30B, and matches the accuracy of the best existing multi-task LLMs (T0-11B and OPT-IML-175B). These findings highlight Cappy’s capabilities and parameter efficiency, which can be credited to its scoring-based pre-training strategy that integrates contrastive information by differentiating between high-quality and low-quality responses. On the contrary, previous multi-task LLMs depend exclusively on teacher-forcing training that utilizes only the ground truth responses.

The overall accuracy averaged over eleven test tasks from PromptSource. “RM” refers to a pre-trained RLHF reward model. Cappy matches the best ones among existing multi-task LLMs.

We also examine the adaptation of multi-task LLMs with Cappy on complex tasks from BIG-Bench, a set of manually curated tasks that are considered beyond the capability of many LLMs. We focus on all the 45 generation BIG-Bench tasks, specifically those that do not offer pre-established answer choices. We evaluate the performance using the Rouge-L score (representing the overall similarity between model generations and corresponding ground truths) on every test set, reporting the average score across 45 tests. In this experiment, all variants of FLAN-T5 serve as the backbone LLMs, and the foundational FLAN-T5 models are frozen. These results, shown below, suggest that Cappy enhances the performance of FLAN-T5 models by a large margin, consistently outperforming the most effective baseline achieved through sample selection using self-scoring of the LLM itself.

The averaged Rouge-L score over 45 complex tasks within BIG-Bench. The x-axis refers to FLAN-T5 models of different sizes. Every dashed line represents an approach working on FLAN-T5s. Self-scoring refers to using the cross-entropy of LLM to select responses. Cappy enhances the performance of FLAN-T5 models by a large margin.

Conclusion

We introduce Cappy, a novel approach that enhances the performance and efficiency of multi-task LLMs. In our experiments, we adapt a single LLM to several domains with Cappy. In the future, Cappy as a pre-trained model can potentially be used in other creative ways beyond on single LLMs.

Acknowledgments

Thanks to Bowen Tan, Jindong Chen, Lei Meng, Abhanshu Sharma and Ewa Dominowska for their valuable feedback. We would also like to thank Eric Xing and Zhiting Hu for their suggestions.

Categories
Offsites

HEAL: A framework for health equity assessment of machine learning performance

Health equity is a major societal concern worldwide with disparities having many causes. These sources include limitations in access to healthcare, differences in clinical treatment, and even fundamental differences in the diagnostic technology. In dermatology for example, skin cancer outcomes are worse for populations such as minorities, those with lower socioeconomic status, or individuals with limited healthcare access. While there is great promise in recent advances in machine learning (ML) and artificial intelligence (AI) to help improve healthcare, this transition from research to bedside must be accompanied by a careful understanding of whether and how they impact health equity.

Health equity is defined by public health organizations as fairness of opportunity for everyone to be as healthy as possible. Importantly, equity may be different from equality. For example, people with greater barriers to improving their health may require more or different effort to experience this fair opportunity. Similarly, equity is not fairness as defined in the AI for healthcare literature. Whereas AI fairness often strives for equal performance of the AI technology across different patient populations, this does not center the goal of prioritizing performance with respect to pre-existing health disparities.

Health equity considerations. An intervention (e.g., an ML-based tool, indicated in dark blue) promotes health equity if it helps reduce existing disparities in health outcomes (indicated in lighter blue).

In “Health Equity Assessment of machine Learning performance (HEAL): a framework and dermatology AI model case study”, published in The Lancet eClinicalMedicine, we propose a methodology to quantitatively assess whether ML-based health technologies perform equitably. In other words, does the ML model perform well for those with the worst health outcomes for the condition(s) the model is meant to address? This goal anchors on the principle that health equity should prioritize and measure model performance with respect to disparate health outcomes, which may be due to a number of factors that include structural inequities (e.g., demographic, social, cultural, political, economic, environmental and geographic).

The health equity framework (HEAL)

The HEAL framework proposes a 4-step process to estimate the likelihood that an ML-based health technology performs equitably:

  1. Identify factors associated with health inequities and define tool performance metrics,
  2. Identify and quantify pre-existing health disparities,
  3. Measure the performance of the tool for each subpopulation,
  4. Measure the likelihood that the tool prioritizes performance with respect to health disparities.

The final step’s output is termed the HEAL metric, which quantifies how anticorrelated the ML model’s performance is with health disparities. In other words, does the model perform better with populations that have the worse health outcomes?

This 4-step process is designed to inform improvements for making ML model performance more equitable, and is meant to be iterative and re-evaluated on a regular basis. For example, the availability of health outcomes data in step (2) can inform the choice of demographic factors and brackets in step (1), and the framework can be applied again with new datasets, models and populations.

Framework for Health Equity Assessment of machine Learning performance (HEAL). Our guiding principle is to avoid exacerbating health inequities, and these steps help us identify disparities and assess for inequitable model performance to move towards better outcomes for all.

With this work, we take a step towards encouraging explicit assessment of the health equity considerations of AI technologies, and encourage prioritization of efforts during model development to reduce health inequities for subpopulations exposed to structural inequities that can precipitate disparate outcomes. We should note that the present framework does not model causal relationships and, therefore, cannot quantify the actual impact a new technology will have on reducing health outcome disparities. However, the HEAL metric may help identify opportunities for improvement, where the current performance is not prioritized with respect to pre-existing health disparities.

Case study on a dermatology model

As an illustrative case study, we applied the framework to a dermatology model, which utilizes a convolutional neural network similar to that described in prior work. This example dermatology model was trained to classify 288 skin conditions using a development dataset of 29k cases. The input to the model consists of three photos of a skin concern along with demographic information and a brief structured medical history. The output consists of a ranked list of possible matching skin conditions.

Using the HEAL framework, we evaluated this model by assessing whether it prioritized performance with respect to pre-existing health outcomes. The model was designed to predict possible dermatologic conditions (from a list of hundreds) based on photos of a skin concern and patient metadata. Evaluation of the model is done using a top-3 agreement metric, which quantifies how often the top 3 output conditions match the most likely condition as suggested by a dermatologist panel. The HEAL metric is computed via the anticorrelation of this top-3 agreement with health outcome rankings.

We used a dataset of 5,420 teledermatology cases, enriched for diversity in age, sex and race/ethnicity, to retrospectively evaluate the model’s HEAL metric. The dataset consisted of “store-and-forward” cases from patients of 20 years or older from primary care providers in the USA and skin cancer clinics in Australia. Based on a review of the literature, we decided to explore race/ethnicity, sex and age as potential factors of inequity, and used sampling techniques to ensure that our evaluation dataset had sufficient representation of all race/ethnicity, sex and age groups. To quantify pre-existing health outcomes for each subgroup we relied on measurements from public databases endorsed by the World Health Organization, such as Years of Life Lost (YLLs) and Disability-Adjusted Life Years (DALYs; years of life lost plus years lived with disability).

HEAL metric for all dermatologic conditions across race/ethnicity subpopulations, including health outcomes (YLLs per 100,000), model performance (top-3 agreement), and rankings for health outcomes and tool performance.
(* Higher is better; measures the likelihood the model performs equitably with respect to the axes in this table.)
HEAL metric for all dermatologic conditions across sexes, including health outcomes (DALYs per 100,000), model performance (top-3 agreement), and rankings for health outcomes and tool performance. (* As above.)

Our analysis estimated that the model was 80.5% likely to perform equitably across race/ethnicity subgroups and 92.1% likely to perform equitably across sexes.

However, while the model was likely to perform equitably across age groups for cancer conditions specifically, we discovered that it had room for improvement across age groups for non-cancer conditions. For example, those 70+ have the poorest health outcomes related to non-cancer skin conditions, yet the model didn’t prioritize performance for this subgroup.

HEAL metrics for all cancer and non-cancer dermatologic conditions across age groups, including health outcomes (DALYs per 100,000), model performance (top-3 agreement), and rankings for health outcomes and tool performance. (* As above.)

Putting things in context

For holistic evaluation, the HEAL metric cannot be employed in isolation. Instead this metric should be contextualized alongside many other factors ranging from computational efficiency and data privacy to ethical values, and aspects that may influence the results (e.g., selection bias or differences in representativeness of the evaluation data across demographic groups).

As an adversarial example, the HEAL metric can be artificially improved by deliberately reducing model performance for the most advantaged subpopulation until performance for that subpopulation is worse than all others. For illustrative purposes, given subpopulations A and B where A has worse health outcomes than B, consider the choice between two models: Model 1 (M1) performs 5% better for subpopulation A than for subpopulation B. Model 2 (M2) performs 5% worse on subpopulation A than B. The HEAL metric would be higher for M1 because it prioritizes performance on a subpopulation with worse outcomes. However, M1 may have absolute performances of just 75% and 70% for subpopulations A and B respectively, while M2 has absolute performances of 75% and 80% for subpopulations A and B respectively. Choosing M1 over M2 would lead to worse overall performance for all subpopulations because some subpopulations are worse-off while no subpopulation is better-off.

Accordingly, the HEAL metric should be used alongside a Pareto condition (discussed further in the paper), which restricts model changes so that outcomes for each subpopulation are either unchanged or improved compared to the status quo, and performance does not worsen for any subpopulation.

The HEAL framework, in its current form, assesses the likelihood that an ML-based model prioritizes performance for subpopulations with respect to pre-existing health disparities for specific subpopulations. This differs from the goal of understanding whether ML will reduce disparities in outcomes across subpopulations in reality. Specifically, modeling improvements in outcomes requires a causal understanding of steps in the care journey that happen both before and after use of any given model. Future research is needed to address this gap.

Conclusion

The HEAL framework enables a quantitative assessment of the likelihood that health AI technologies prioritize performance with respect to health disparities. The case study demonstrates how to apply the framework in the dermatological domain, indicating a high likelihood that model performance is prioritized with respect to health disparities across sex and race/ethnicity, but also revealing the potential for improvements for non-cancer conditions across age. The case study also illustrates limitations in the ability to apply all recommended aspects of the framework (e.g., mapping societal context, availability of data), thus highlighting the complexity of health equity considerations of ML-based tools.

This work is a proposed approach to address a grand challenge for AI and health equity, and may provide a useful evaluation framework not only during model development, but during pre-implementation and real-world monitoring stages, e.g., in the form of health equity dashboards. We hold that the strength of the HEAL framework is in its future application to various AI tools and use cases and its refinement in the process. Finally, we acknowledge that a successful approach towards understanding the impact of AI technologies on health equity needs to be more than a set of metrics. It will require a set of goals agreed upon by a community that represents those who will be most impacted by a model.

Acknowledgements

The research described here is joint work across many teams at Google. We are grateful to all our co-authors: Terry Spitz, Malcolm Pyles, Heather Cole-Lewis, Ellery Wulczyn, Stephen R. Pfohl, Donald Martin, Jr., Ronnachai Jaroensri, Geoff Keeling, Yuan Liu, Stephanie Farquhar, Qinghan Xue, Jenna Lester, Cían Hughes, Patricia Strachan, Fraser Tan, Peggy Bui, Craig H. Mermel, Lily H. Peng, Yossi Matias, Greg S. Corrado, Dale R. Webster, Sunny Virmani, Christopher Semturs, Yun Liu, and Po-Hsuan Cameron Chen. We also thank Lauren Winer, Sami Lachgar, Ting-An Lin, Aaron Loh, Morgan Du, Jenny Rizk, Renee Wong, Ashley Carrick, Preeti Singh, Annisah Um’rani, Jessica Schrouff, Alexander Brown, and Anna Iurchenko for their support of this project.

Categories
Offsites

MELON: Reconstructing 3D objects from images with unknown poses

A person’s prior experience and understanding of the world generally enables them to easily infer what an object looks like in whole, even if only looking at a few 2D pictures of it. Yet the capacity for a computer to reconstruct the shape of an object in 3D given only a few images has remained a difficult algorithmic problem for years. This fundamental computer vision task has applications ranging from the creation of e-commerce 3D models to autonomous vehicle navigation.

A key part of the problem is how to determine the exact positions from which images were taken, known as pose inference. If camera poses are known, a range of successful techniques — such as neural radiance fields (NeRF) or 3D Gaussian Splatting — can reconstruct an object in 3D. But if these poses are not available, then we face a difficult “chicken and egg” problem where we could determine the poses if we knew the 3D object, but we can’t reconstruct the 3D object until we know the camera poses. The problem is made harder by pseudo-symmetries — i.e., many objects look similar when viewed from different angles. For example, square objects like a chair tend to look similar every 90° rotation. Pseudo-symmetries of an object can be revealed by rendering it on a turntable from various angles and plotting its photometric self-similarity map.

Self-Similarity map of a toy truck model. Left: The model is rendered on a turntable from various azimuthal angles, θ. Right: The average L2 RGB similarity of a rendering from θ with that of θ*. The pseudo-similarities are indicated by the dashed red lines.

The diagram above only visualizes one dimension of rotation. It becomes even more complex (and difficult to visualize) when introducing more degrees of freedom. Pseudo-symmetries make the problem ill-posed, with naïve approaches often converging to local minima. In practice, such an approach might mistake the back view as the front view of an object, because they share a similar silhouette. Previous techniques (such as BARF or SAMURAI) side-step this problem by relying on an initial pose estimate that starts close to the global minima. But how can we approach this if those aren’t available?

Methods, such as GNeRF and VMRF leverage generative adversarial networks (GANs) to overcome the problem. These techniques have the ability to artificially “amplify” a limited number of training views, aiding reconstruction. GAN techniques, however, often have complex, sometimes unstable, training processes, making robust and reliable convergence difficult to achieve in practice. A range of other successful methods, such as SparsePose or RUST, can infer poses from a limited number views, but require pre-training on a large dataset of posed images, which aren’t always available, and can suffer from “domain-gap” issues when inferring poses for different types of images.

In “MELON: NeRF with Unposed Images in SO(3)”, spotlighted at 3DV 2024, we present a technique that can determine object-centric camera poses entirely from scratch while reconstructing the object in 3D. MELON (Modulo Equivalent Latent Optimization of NeRF) is one of the first techniques that can do this without initial pose camera estimates, complex training schemes or pre-training on labeled data. MELON is a relatively simple technique that can easily be integrated into existing NeRF methods. We demonstrate that MELON can reconstruct a NeRF from unposed images with state-of-the-art accuracy while requiring as few as 4–6 images of an object.

MELON

We leverage two key techniques to aid convergence of this ill-posed problem. The first is a very lightweight, dynamically trained convolutional neural network (CNN) encoder that regresses camera poses from training images. We pass a downscaled training image to a four layer CNN that infers the camera pose. This CNN is initialized from noise and requires no pre-training. Its capacity is so small that it forces similar looking images to similar poses, providing an implicit regularization greatly aiding convergence.

The second technique is a modulo loss that simultaneously considers pseudo symmetries of an object. We render the object from a fixed set of viewpoints for each training image, backpropagating the loss only through the view that best fits the training image. This effectively considers the plausibility of multiple views for each image. In practice, we find N=2 views (viewing an object from the other side) is all that’s required in most cases, but sometimes get better results with N=4 for square objects.

These two techniques are integrated into standard NeRF training, except that instead of fixed camera poses, poses are inferred by the CNN and duplicated by the modulo loss. Photometric gradients back-propagate through the best-fitting cameras into the CNN. We observe that cameras generally converge quickly to globally optimal poses (see animation below). After training of the neural field, MELON can synthesize novel views using standard NeRF rendering methods.

We simplify the problem by using the NeRF-Synthetic dataset, a popular benchmark for NeRF research and common in the pose-inference literature. This synthetic dataset has cameras at precisely fixed distances and a consistent “up” orientation, requiring us to infer only the polar coordinates of the camera. This is the same as an object at the center of a globe with a camera always pointing at it, moving along the surface. We then only need the latitude and longitude (2 degrees of freedom) to specify the camera pose.

MELON uses a dynamically trained lightweight CNN encoder that predicts a pose for each image. Predicted poses are replicated by the modulo loss, which only penalizes the smallest L2 distance from the ground truth color. At evaluation time, the neural field can be used to generate novel views.

Results

We compute two key metrics to evaluate MELON’s performance on the NeRF Synthetic dataset. The error in orientation between the ground truth and inferred poses can be quantified as a single angular error that we average across all training images, the pose error. We then test the accuracy of MELON’s rendered objects from novel views by measuring the peak signal-to-noise ratio (PSNR) against held out test views. We see that MELON quickly converges to the approximate poses of most cameras within the first 1,000 steps of training, and achieves a competitive PSNR of 27.5 dB after 50k steps.

Convergence of MELON on a toy truck model during optimization. Left: Rendering of the NeRF. Right: Polar plot of predicted (blue x), and ground truth (red dot) cameras.

MELON achieves similar results for other scenes in the NeRF Synthetic dataset.

Reconstruction quality comparison between ground-truth (GT) and MELON on NeRF-Synthetic scenes after 100k training steps.

Noisy images

MELON also works well when performing novel view synthesis from extremely noisy, unposed images. We add varying amounts, σ, of white Gaussian noise to the training images. For example, the object in σ=1.0 below is impossible to make out, yet MELON can determine the pose and generate novel views of the object.

Novel view synthesis from noisy unposed 128×128 images. Top: Example of noise level present in training views. Bottom: Reconstructed model from noisy training views and mean angular pose error.

This perhaps shouldn’t be too surprising, given that techniques like RawNeRF have demonstrated NeRF’s excellent de-noising capabilities with known camera poses. The fact that MELON works for noisy images of unknown camera poses so robustly was unexpected.

Conclusion

We present MELON, a technique that can determine object-centric camera poses to reconstruct objects in 3D without the need for approximate pose initializations, complex GAN training schemes or pre-training on labeled data. MELON is a relatively simple technique that can easily be integrated into existing NeRF methods. Though we only demonstrated MELON on synthetic images we are adapting our technique to work in real world conditions. See the paper and MELON site to learn more.

Acknowledgements

We would like to thank our paper co-authors Axel Levy, Matan Sela, and Gordon Wetzstein, as well as Florian Schroff and Hartwig Adam for continuous help in building this technology. We also thank Matthew Brown, Ricardo Martin-Brualla and Frederic Poitevin for their helpful feedback on the paper draft. We also acknowledge the use of the computational resources at the SLAC Shared Scientific Data Facility (SDF).

Categories
Offsites

SCIN: A new resource for representative dermatology images

Health datasets play a crucial role in research and medical education, but it can be challenging to create a dataset that represents the real world. For example, dermatology conditions are diverse in their appearance and severity and manifest differently across skin tones. Yet, existing dermatology image datasets often lack representation of everyday conditions (like rashes, allergies and infections) and skew towards lighter skin tones. Furthermore, race and ethnicity information is frequently missing, hindering our ability to assess disparities or create solutions.

To address these limitations, we are releasing the Skin Condition Image Network (SCIN) dataset in collaboration with physicians at Stanford Medicine. We designed SCIN to reflect the broad range of concerns that people search for online, supplementing the types of conditions typically found in clinical datasets. It contains images across various skin tones and body parts, helping to ensure that future AI tools work effectively for all. We’ve made the SCIN dataset freely available as an open-access resource for researchers, educators, and developers, and have taken careful steps to protect contributor privacy.

Example set of images and metadata from the SCIN dataset.

Dataset composition

The SCIN dataset currently contains over 10,000 images of skin, nail, or hair conditions, directly contributed by individuals experiencing them. All contributions were made voluntarily with informed consent by individuals in the US, under an institutional-review board approved study. To provide context for retrospective dermatologist labeling, contributors were asked to take images both close-up and from slightly further away. They were given the option to self-report demographic information and tanning propensity (self-reported Fitzpatrick Skin Type, i.e., sFST), and to describe the texture, duration and symptoms related to their concern.

One to three dermatologists labeled each contribution with up to five dermatology conditions, along with a confidence score for each label. The SCIN dataset contains these individual labels, as well as an aggregated and weighted differential diagnosis derived from them that could be useful for model testing or training. These labels were assigned retrospectively and are not equivalent to a clinical diagnosis, but they allow us to compare the distribution of dermatology conditions in the SCIN dataset with existing datasets.

The SCIN dataset contains largely allergic, inflammatory and infectious conditions while datasets from clinical sources focus on benign and malignant neoplasms.

While many existing dermatology datasets focus on malignant and benign tumors and are intended to assist with skin cancer diagnosis, the SCIN dataset consists largely of common allergic, inflammatory, and infectious conditions. The majority of images in the SCIN dataset show early-stage concerns — more than half arose less than a week before the photo, and 30% arose less than a day before the image was taken. Conditions within this time window are seldom seen within the health system and therefore are underrepresented in existing dermatology datasets.

We also obtained dermatologist estimates of Fitzpatrick Skin Type (estimated FST or eFST) and layperson labeler estimates of Monk Skin Tone (eMST) for the images. This allowed comparison of the skin condition and skin type distributions to those in existing dermatology datasets. Although we did not selectively target any skin types or skin tones, the SCIN dataset has a balanced Fitzpatrick skin type distribution (with more of Types 3, 4, 5, and 6) compared to similar datasets from clinical sources.

Self-reported and dermatologist-estimated Fitzpatrick Skin Type distribution in the SCIN dataset compared with existing un-enriched dermatology datasets (Fitzpatrick17k, PH², SKINL2, and PAD-UFES-20).

The Fitzpatrick Skin Type scale was originally developed as a photo-typing scale to measure the response of skin types to UV radiation, and it is widely used in dermatology research. The Monk Skin Tone scale is a newer 10-shade scale that measures skin tone rather than skin phototype, capturing more nuanced differences between the darker skin tones. While neither scale was intended for retrospective estimation using images, the inclusion of these labels is intended to enable future research into skin type and tone representation in dermatology. For example, the SCIN dataset provides an initial benchmark for the distribution of these skin types and tones in the US population.

The SCIN dataset has a high representation of women and younger individuals, likely reflecting a combination of factors. These could include differences in skin condition incidence, propensity to seek health information online, and variations in willingness to contribute to research across demographics.

Crowdsourcing method

To create the SCIN dataset, we used a novel crowdsourcing method, which we describe in the accompanying research paper co-authored with investigators at Stanford Medicine. This approach empowers individuals to play an active role in healthcare research. It allows us to reach people at earlier stages of their health concerns, potentially before they seek formal care. Crucially, this method uses advertisements on web search result pages — the starting point for many people’s health journey — to connect with participants.

Our results demonstrate that crowdsourcing can yield a high-quality dataset with a low spam rate. Over 97.5% of contributions were genuine images of skin conditions. After performing further filtering steps to exclude images that were out of scope for the SCIN dataset and to remove duplicates, we were able to release nearly 90% of the contributions received over the 8-month study period. Most images were sharp and well-exposed. Approximately half of the contributions include self-reported demographics, and 80% contain self-reported information relating to the skin condition, such as texture, duration, or other symptoms. We found that dermatologists’ ability to retrospectively assign a differential diagnosis depended more on the availability of self-reported information than on image quality.

Dermatologist confidence in their labels (scale from 1-5) depended on the availability of self-reported demographic and symptom information.

While perfect image de-identification can never be guaranteed, protecting the privacy of individuals who contributed their images was a top priority when creating the SCIN dataset. Through informed consent, contributors were made aware of potential re-identification risks and advised to avoid uploading images with identifying features. Post-submission privacy protection measures included manual redaction or cropping to exclude potentially identifying areas, reverse image searches to exclude publicly available copies and metadata removal or aggregation. The SCIN Data Use License prohibits attempts to re-identify contributors.

We hope the SCIN dataset will be a helpful resource for those working to advance inclusive dermatology research, education, and AI tool development. By demonstrating an alternative to traditional dataset creation methods, SCIN paves the way for more representative datasets in areas where self-reported data or retrospective labeling is feasible.

Acknowledgements

We are grateful to all our co-authors Abbi Ward, Jimmy Li, Julie Wang, Sriram Lakshminarasimhan, Ashley Carrick, Bilson Campana, Jay Hartford, Pradeep Kumar S, Tiya Tiyasirisokchai, Sunny Virmani, Renee Wong, Yossi Matias, Greg S. Corrado, Dale R. Webster, Dawn Siegel (Stanford Medicine), Steven Lin (Stanford Medicine), Justin Ko (Stanford Medicine), Alan Karthikesalingam and Christopher Semturs. We also thank Yetunde Ibitoye, Sami Lachgar, Lisa Lehmann, Javier Perez, Margaret Ann Smith (Stanford Medicine), Rachelle Sico, Amit Talreja, Annisah Um’rani and Wayne Westerlind for their essential contributions to this work. Finally, we are grateful to Heather Cole-Lewis, Naama Hammel, Ivor Horn, Michael Howell, Yun Liu, and Eric Teasley for their insightful comments on the study design and manuscript.

Categories
Offsites

ScreenAI: A visual language model for UI and visually-situated language understanding

Screen user interfaces (UIs) and infographics, such as charts, diagrams and tables, play important roles in human communication and human-machine interaction as they facilitate rich and interactive user experiences. UIs and infographics share similar design principles and visual language (e.g., icons and layouts), that offer an opportunity to build a single model that can understand, reason, and interact with these interfaces. However, because of their complexity and varied presentation formats, infographics and UIs present a unique modeling challenge.

To that end, we introduce “ScreenAI: A Vision-Language Model for UI and Infographics Understanding”. ScreenAI improves upon the PaLI architecture with the flexible patching strategy from pix2struct. We train ScreenAI on a unique mixture of datasets and tasks, including a novel Screen Annotation task that requires the model to identify UI element information (i.e., type, location and description) on a screen. These text annotations provide large language models (LLMs) with screen descriptions, enabling them to automatically generate question-answering (QA), UI navigation, and summarization training datasets at scale. At only 5B parameters, ScreenAI achieves state-of-the-art results on UI- and infographic-based tasks (WebSRC and MoTIF), and best-in-class performance on Chart QA, DocVQA, and InfographicVQA compared to models of similar size. We are also releasing three new datasets: Screen Annotation to evaluate the layout understanding capability of the model, as well as ScreenQA Short and Complex ScreenQA for a more comprehensive evaluation of its QA capability.

ScreenAI

ScreenAI’s architecture is based on PaLI, composed of a multimodal encoder block and an autoregressive decoder. The PaLI encoder uses a vision transformer (ViT) that creates image embeddings and a multimodal encoder that takes the concatenation of the image and text embeddings as input. This flexible architecture allows ScreenAI to solve vision tasks that can be recast as text+image-to-text problems.

On top of the PaLI architecture, we employ a flexible patching strategy introduced in pix2struct. Instead of using a fixed-grid pattern, the grid dimensions are selected such that they preserve the native aspect ratio of the input image. This enables ScreenAI to work well across images of various aspect ratios.

The ScreenAI model is trained in two stages: a pre-training stage followed by a fine-tuning stage. First, self-supervised learning is applied to automatically generate data labels, which are then used to train ViT and the language model. ViT is frozen during the fine-tuning stage, where most data used is manually labeled by human raters.

ScreenAI model architecture.

Data generation

To create a pre-training dataset for ScreenAI, we first compile an extensive collection of screenshots from various devices, including desktops, mobile, and tablets. This is achieved by using publicly accessible web pages and following the programmatic exploration approach used for the RICO dataset for mobile apps. We then apply a layout annotator, based on the DETR model, that identifies and labels a wide range of UI elements (e.g., image, pictogram, button, text) and their spatial relationships. Pictograms undergo further analysis using an icon classifier capable of distinguishing 77 different icon types. This detailed classification is essential for interpreting the subtle information conveyed through icons. For icons that are not covered by the classifier, and for infographics and images, we use the PaLI image captioning model to generate descriptive captions that provide contextual information. We also apply an optical character recognition (OCR) engine to extract and annotate textual content on screen. We combine the OCR text with the previous annotations to create a detailed description of each screen.

A mobile app screenshot with generated annotations that include UI elements and their descriptions, e.g., TEXT elements also contain the text content from OCR, IMAGE elements contain image captions, LIST_ITEMs contain all their child elements.

LLM-based data generation

We enhance the pre-training data’s diversity using PaLM 2 to generate input-output pairs in a two-step process. First, screen annotations are generated using the technique outlined above, then we craft a prompt around this schema for the LLM to create synthetic data. This process requires prompt engineering and iterative refinement to find an effective prompt. We assess the generated data’s quality through human validation against a quality threshold.

You only speak JSON. Do not write text that isn’t JSON.
You are given the following mobile screenshot, described in words. Can you generate 5 questions regarding the content of the screenshot as well as the corresponding short answers to them? 

The answer should be as short as possible, containing only the necessary information. Your answer should be structured as follows:
questions: [
{{question: the question,
    answer: the answer
}},
 ...
]

{THE SCREEN SCHEMA}

A sample prompt for QA data generation.

By combining the natural language capabilities of LLMs with a structured schema, we simulate a wide range of user interactions and scenarios to generate synthetic, realistic tasks. In particular, we generate three categories of tasks:

  • Question answering: The model is asked to answer questions regarding the content of the screenshots, e.g., “When does the restaurant open?”
  • Screen navigation: The model is asked to convert a natural language utterance into an executable action on a screen, e.g., “Click the search button.”
  • Screen summarization: The model is asked to summarize the screen content in one or two sentences.
Block diagram of our workflow for generating data for QA, summarization and navigation tasks using existing ScreenAI models and LLMs. Each task uses a custom prompt to emphasize desired aspects, like questions related to counting, involving reasoning, etc.

LLM-generated data. Examples for screen QA, navigation and summarization. For navigation, the action bounding box is displayed in red on the screenshot.

Experiments and results

As previously mentioned, ScreenAI is trained in two stages: pre-training and fine-tuning. Pre-training data labels are obtained using self-supervised learning and fine-tuning data labels comes from human raters.

We fine-tune ScreenAI using public QA, summarization, and navigation datasets and a variety of tasks related to UIs. For QA, we use well established benchmarks in the multimodal and document understanding field, such as ChartQA, DocVQA, Multi page DocVQA, InfographicVQA, OCR VQA, Web SRC and ScreenQA. For navigation, datasets used include Referring Expressions, MoTIF, Mug, and Android in the Wild. Finally, we use Screen2Words for screen summarization and Widget Captioning for describing specific UI elements. Along with the fine-tuning datasets, we evaluate the fine-tuned ScreenAI model using three novel benchmarks:

  1. Screen Annotation: Enables the evaluation model layout annotations and spatial understanding capabilities.
  2. ScreenQA Short: A variation of ScreenQA, where its ground truth answers have been shortened to contain only the relevant information that better aligns with other QA tasks.
  3. Complex ScreenQA: Complements ScreenQA Short with more difficult questions (counting, arithmetic, comparison, and non-answerable questions) and contains screens with various aspect ratios.

The fine-tuned ScreenAI model achieves state-of-the-art results on various UI and infographic-based tasks (WebSRC and MoTIF) and best-in-class performance on Chart QA, DocVQA, and InfographicVQA compared to models of similar size. ScreenAI achieves competitive performance on Screen2Words and OCR-VQA. Additionally, we report results on the new benchmark datasets introduced to serve as a baseline for further research.

Comparing model performance of ScreenAI with state-of-the-art (SOTA) models of similar size.

Next, we examine ScreenAI’s scaling capabilities and observe that across all tasks, increasing the model size improves performances and the improvements have not saturated at the largest size.

Model performance increases with size, and the performance has not saturated even at the largest size of 5B params.

Conclusion

We introduce the ScreenAI model along with a unified representation that enables us to develop self-supervised learning tasks leveraging data from all these domains. We also illustrate the impact of data generation using LLMs and investigate improving model performance on specific aspects with modifying the training mixture. We apply all of these techniques to build multi-task trained models that perform competitively with state-of-the-art approaches on a number of public benchmarks. However, we also note that our approach still lags behind large models and further research is needed to bridge this gap.

Acknowledgements

This project is the result of joint work with Maria Wang, Fedir Zubach, Hassan Mansoor, Vincent Etter, Victor Carbune, Jason Lin, Jindong Chen and Abhanshu Sharma. We thank Fangyu Liu, Xi Chen, Efi Kokiopoulou, Jesse Berent, Gabriel Barcik, Lukas Zilka, Oriana Riva, Gang Li,Yang Li, Radu Soricut, and Tania Bedrax-Weiss for their insightful feedback and discussions, along with Rahul Aralikatte, Hao Cheng and Daniel Kim for their support in data preparation. We also thank Jay Yagnik, Blaise Aguera y Arcas, Ewa Dominowska, David Petrou, and Matt Sharifi for their leadership, vision and support. We are very grateful toTom Small for helping us create the animation in this post.

Categories
Offsites

Using AI to expand global access to reliable flood forecasts

Floods are the most common natural disaster, and are responsible for roughly $50 billion in annual financial damages worldwide. The rate of flood-related disasters has more than doubled since the year 2000 partly due to climate change. Nearly 1.5 billion people, making up 19% of the world’s population, are exposed to substantial risks from severe flood events. Upgrading early warning systems to make accurate and timely information accessible to these populations can save thousands of lives per year.

Driven by the potential impact of reliable flood forecasting on people’s lives globally, we started our flood forecasting effort in 2017. Through this multi-year journey, we advanced research over the years hand-in-hand with building a real-time operational flood forecasting system that provides alerts on Google Search, Maps, Android notifications and through the Flood Hub. However, in order to scale globally, especially in places where accurate local data is not available, more research advances were required.

In “Global prediction of extreme floods in ungauged watersheds”, published in Nature, we demonstrate how machine learning (ML) technologies can significantly improve global-scale flood forecasting relative to the current state-of-the-art for countries where flood-related data is scarce. With these AI-based technologies we extended the reliability of currently-available global nowcasts, on average, from zero to five days, and improved forecasts across regions in Africa and Asia to be similar to what are currently available in Europe. The evaluation of the models was conducted in collaboration with the European Center for Medium Range Weather Forecasting (ECMWF).

These technologies also enable Flood Hub to provide real-time river forecasts up to seven days in advance, covering river reaches across over 80 countries. This information can be used by people, communities, governments and international organizations to take anticipatory action to help protect vulnerable populations.

Flood forecasting at Google

The ML models that power the FloodHub tool are the product of many years of research, conducted in collaboration with several partners, including academics, governments, international organizations, and NGOs.

In 2018, we launched a pilot early warning system in the Ganges-Brahmaputra river basin in India, with the hypothesis that ML could help address the challenging problem of reliable flood forecasting at scale. The pilot was further expanded the following year via the combination of an inundation model, real-time water level measurements, the creation of an elevation map and hydrologic modeling.

In collaboration with academics, and, in particular, with the JKU Institute for Machine Learning we explored ML-based hydrologic models, showing that LSTM-based models could produce more accurate simulations than traditional conceptual and physics-based hydrology models. This research led to flood forecasting improvements that enabled the expansion of our forecasting coverage to include all of India and Bangladesh. We also worked with researchers at Yale University to test technological interventions that increase the reach and impact of flood warnings.

Our hydrological models predict river floods by processing publicly available weather data like precipitation and physical watershed information. Such models must be calibrated to long data records from streamflow gauging stations in individual rivers. A low percentage of global river watersheds (basins) have streamflow gauges, which are expensive but necessary to supply relevant data, and it’s challenging for hydrological simulation and forecasting to provide predictions in basins that lack this infrastructure. Lower gross domestic product (GDP) is correlated with increased vulnerability to flood risks, and there is an inverse correlation between national GDP and the amount of publicly available data in a country. ML helps to address this problem by allowing a single model to be trained on all available river data and to be applied to ungauged basins where no data are available. In this way, models can be trained globally, and can make predictions for any river location.

There is an inverse (log-log) correlation between the amount of publicly available streamflow data in a country and national GDP. Streamflow data from the Global Runoff Data Center.

Our academic collaborations led to ML research that developed methods to estimate uncertainty in river forecasts and showed how ML river forecast models synthesize information from multiple data sources. They demonstrated that these models can simulate extreme events reliably, even when those events are not part of the training data. In an effort to contribute to open science, in 2023 we open-sourced a community-driven dataset for large-sample hydrology in Nature Scientific Data.

The river forecast model

Most hydrology models used by national and international agencies for flood forecasting and river modeling are state-space models, which depend only on daily inputs (e.g., precipitation, temperature, etc.) and the current state of the system (e.g., soil moisture, snowpack, etc.). LSTMs are a variant of state-space models and work by defining a neural network that represents a single time step, where input data (such as current weather conditions) are processed to produce updated state information and output values (streamflow) for that time step. LSTMs are applied sequentially to make time-series predictions, and in this sense, behave similarly to how scientists typically conceptualize hydrologic systems. Empirically, we have found that LSTMs perform well on the task of river forecasting.

A diagram of the LSTM, which is a neural network that operates sequentially in time. An accessible primer can be found here.

Our river forecast model uses two LSTMs applied sequentially: (1) a “hindcast” LSTM ingests historical weather data (dynamic hindcast features) up to the present time (or rather, the issue time of a forecast), and (2) a “forecast” LSTM ingests states from the hindcast LSTM along with forecasted weather data (dynamic forecast features) to make future predictions. One year of historical weather data are input into the hindcast LSTM, and seven days of forecasted weather data are input into the forecast LSTM. Static features include geographical and geophysical characteristics of watersheds that are input into both the hindcast and forecast LSTMs and allow the model to learn different hydrological behaviors and responses in various types of watersheds.

Output from the forecast LSTM is fed into a “head” layer that uses mixture density networks to produce a probabilistic forecast (i.e., predicted parameters of a probability distribution over streamflow). Specifically, the model predicts the parameters of a mixture of heavy-tailed probability density functions, called asymmetric Laplacian distributions, at each forecast time step. The result is a mixture density function, called a Countable Mixture of Asymmetric Laplacians (CMAL) distribution, which represents a probabilistic prediction of the volumetric flow rate in a particular river at a particular time.

LSTM-based river forecast model architecture. Two LSTMs are applied in sequence, one ingesting historical weather data and one ingesting forecasted weather data. The model outputs are the parameters of a probability distribution over streamflow at each forecasted timestep.

Input and training data

The model uses three types of publicly available data inputs, mostly from governmental sources:

  1. Static watershed attributes representing geographical and geophysical variables: From the HydroATLAS project, including data like long-term climate indexes (precipitation, temperature, snow fractions), land cover, and anthropogenic attributes (e.g., a nighttime lights index as a proxy for human development).
  2. Historical meteorological time-series data: Used to spin up the model for one year prior to the issue time of a forecast. The data comes from NASA IMERG, NOAA CPC Global Unified Gauge-Based Analysis of Daily Precipitation, and the ECMWF ERA5-land reanalysis. Variables include daily total precipitation, air temperature, solar and thermal radiation, snowfall, and surface pressure.
  3. Forecasted meteorological time series over a seven-day forecast horizon: Used as input for the forecast LSTM. These data are the same meteorological variables listed above, and come from the ECMWF HRES atmospheric model.

Training data are daily streamflow values from the Global Runoff Data Center over the time period 1980 – 2023. A single streamflow forecast model is trained using data from 5,680 diverse watershed streamflow gauges (shown below) to improve accuracy.

Location of 5,680 streamflow gauges that supply training data for the river forecast model from the Global Runoff Data Center.

Improving on the current state-of-the-art

We compared our river forecast model with GloFAS version 4, the current state-of-the-art global flood forecasting system. These experiments showed that ML can provide accurate warnings earlier and over larger and more impactful events.

The figure below shows the distribution of F1 scores when predicting different severity events at river locations around the world, with plus or minus 1 day accuracy. F1 scores are an average of precision and recall and event severity is measured by return period. For example, a 2-year return period event is a volume of streamflow that is expected to be exceeded on average once every two years. Our model achieves reliability scores at up to 4-day or 5-day lead times that are similar to or better, on average, than the reliability of GloFAS nowcasts (0-day lead time).

Distributions of F1 scores over 2-year return period events in 2,092 watersheds globally during the time period 2014-2023 from GloFAS (blue) and our model (orange) at different lead times. On average, our model is statistically as accurate as GloFAS nowcasts (0–day lead time) up to 5 days in advance over 2-year (shown) and 1-year, 5-year, and 10-year events (not shown).

Additionally (not shown), our model achieves accuracies over larger and rarer extreme events, with precision and recall scores over 5-year return period events that are similar to or better than GloFAS accuracies over 1-year return period events. See the paper for more information.

Looking into the future

The flood forecasting initiative is part of our Adaptation and Resilience efforts and reflects Google’s commitment to address climate change while helping global communities become more resilient. We believe that AI and ML will continue to play a critical role in helping advance science and research towards climate action.

We actively collaborate with several international aid organizations (e.g., the Centre for Humanitarian Data and the Red Cross) to provide actionable flood forecasts. Additionally, in an ongoing collaboration with the World Meteorological Organization (WMO) to support early warning systems for climate hazards, we are conducting a study to help understand how AI can help address real-world challenges faced by national flood forecasting agencies.

While the work presented here demonstrates a significant step forward in flood forecasting, future work is needed to further expand flood forecasting coverage to more locations globally and other types of flood-related events and disasters, including flash floods and urban floods. We are looking forward to continuing collaborations with our partners in the academic and expert communities, local governments and the industry to reach these goals.

Categories
Offsites

Computer-aided diagnosis for lung cancer screening

Lung cancer is the leading cause of cancer-related deaths globally with 1.8 million deaths reported in 2020. Late diagnosis dramatically reduces the chances of survival. Lung cancer screening via computed tomography (CT), which provides a detailed 3D image of the lungs, has been shown to reduce mortality in high-risk populations by at least 20% by detecting potential signs of cancers earlier. In the US, screening involves annual scans, with some countries or cases recommending more or less frequent scans.

The United States Preventive Services Task Force recently expanded lung cancer screening recommendations by roughly 80%, which is expected to increase screening access for women and racial and ethnic minority groups. However, false positives (i.e., incorrectly reporting a potential cancer in a cancer-free patient) can cause anxiety and lead to unnecessary procedures for patients while increasing costs for the healthcare system. Moreover, efficiency in screening a large number of individuals can be challenging depending on healthcare infrastructure and radiologist availability.

At Google we have previously developed machine learning (ML) models for lung cancer detection, and have evaluated their ability to automatically detect and classify regions that show signs of potential cancer. Performance has been shown to be comparable to that of specialists in detecting possible cancer. While they have achieved high performance, effectively communicating findings in realistic environments is necessary to realize their full potential.

To that end, in “Assistive AI in Lung Cancer Screening: A Retrospective Multinational Study in the US and Japan”, published in Radiology AI, we investigate how ML models can effectively communicate findings to radiologists. We also introduce a generalizable user-centric interface to help radiologists leverage such models for lung cancer screening. The system takes CT imaging as input and outputs a cancer suspicion rating using four categories (no suspicion, probably benign, suspicious, highly suspicious) along with the corresponding regions of interest. We evaluate the system’s utility in improving clinician performance through randomized reader studies in both the US and Japan, using the local cancer scoring systems (Lung-RADSs V1.1 and Sendai Score) and image viewers that mimic realistic settings. We found that reader specificity increases with model assistance in both reader studies. To accelerate progress in conducting similar studies with ML models, we have open-sourced code to process CT images and generate images compatible with the picture archiving and communication system (PACS) used by radiologists.

Developing an interface to communicate model results

Integrating ML models into radiologist workflows involves understanding the nuances and goals of their tasks to meaningfully support them. In the case of lung cancer screening, hospitals follow various country-specific guidelines that are regularly updated. For example, in the US, Lung-RADs V1.1 assigns an alpha-numeric score to indicate the lung cancer risk and follow-up recommendations. When assessing patients, radiologists load the CT in their workstation to read the case, find lung nodules or lesions, and apply set guidelines to determine follow-up decisions.

Our first step was to improve the previously developed ML models through additional training data and architectural improvements, including self-attention. Then, instead of targeting specific guidelines, we experimented with a complementary way of communicating AI results independent of guidelines or their particular versions. Specifically, the system output offers a suspicion rating and localization (regions of interest) for the user to consider in conjunction with their own specific guidelines. The interface produces output images directly associated with the CT study, requiring no changes to the user’s workstation. The radiologist only needs to review a small set of additional images. There is no other change to their system or interaction with the system.

Example of the assistive lung cancer screening system outputs. Results for the radiologist’s evaluation are visualized on the location of the CT volume where the suspicious lesion is found. The overall suspicion is displayed at the top of the CT images. Circles highlight the suspicious lesions while squares show a rendering of the same lesion from a different perspective, called a sagittal view.

The assistive lung cancer screening system comprises 13 models and has a high-level architecture similar to the end-to-end system used in prior work. The models coordinate with each other to first segment the lungs, obtain an overall assessment, locate three suspicious regions, then use the information to assign a suspicion rating to each region. The system was deployed on Google Cloud using a Google Kubernetes Engine (GKE) that pulled the images, ran the ML models, and provided results. This allows scalability and directly connects to servers where the images are stored in DICOM stores.

Outline of the Google Cloud deployment of the assistive lung cancer screening system and the directional calling flow for the individual components that serve the images and compute results. Images are served to the viewer and to the system using Google Cloud services. The system is run on a Google Kubernetes Engine that pulls the images, processes them, and writes them back into the DICOM store.

Reader studies

To evaluate the system’s utility in improving clinical performance, we conducted two reader studies (i.e., experiments designed to assess clinical performance comparing expert performance with and without the aid of a technology) with 12 radiologists using pre-existing, de-identified CT scans. We presented 627 challenging cases to 6 US-based and 6 Japan-based radiologists. In the experimental setup, readers were divided into two groups that read each case twice, with and without assistance from the model. Readers were asked to apply scoring guidelines they typically use in their clinical practice and report their overall suspicion of cancer for each case. We then compared the results of the reader’s responses to measure the impact of the model on their workflow and decisions. The score and suspicion level were judged against the actual cancer outcomes of the individuals to measure sensitivity, specificity, and area under the ROC curve (AUC) values. These were compared with and without assistance.

A multi-case multi-reader study involves each case being reviewed by each reader twice, once with ML system assistance and once without. In this visualization one reader first reviews Set A without assistance (blue) and then with assistance (orange) after a wash-out period. A second reader group follows the opposite path by reading the same set of cases Set A with assistance first. Readers are randomized to these groups to remove the effect of ordering.

The ability to conduct these studies using the same interface highlights its generalizability to completely different cancer scoring systems, and the generalization of the model and assistive capability to different patient populations. Our study results demonstrated that when radiologists used the system in their clinical evaluation, they had an increased ability to correctly identify lung images without actionable lung cancer findings (i.e., specificity) by an absolute 5–7% compared to when they didn’t use the assistive system. This potentially means that for every 15–20 patients screened, one may be able to avoid unnecessary follow-up procedures, thus reducing their anxiety and the burden on the health care system. This can, in turn, help improve the sustainability of lung cancer screening programs, particularly as more people become eligible for screening.

Reader specificity increases with ML model assistance in both the US-based and Japan-based reader studies. Specificity values were derived from reader scores from actionable findings (something suspicious was found) versus no actionable findings, compared against the true cancer outcome of the individual. Under model assistance, readers flagged fewer cancer-negative individuals for follow-up visits. Sensitivity for cancer positive individuals remained the same.

Translating this into real-world impact through partnership

The system results demonstrate the potential for fewer follow-up visits, reduced anxiety, as well lower overall costs for lung cancer screening. In an effort to translate this research into real-world clinical impact, we are working with: DeepHealth, a leading AI-powered health informatics provider; and Apollo Radiology International a leading provider of Radiology services in India to explore paths for incorporating this system into future products. In addition, we are looking to help other researchers studying how best to integrate ML model results into clinical workflows by open sourcing code used for the reader study and incorporating the insights described in this blog. We hope that this will help accelerate medical imaging researchers looking to conduct reader studies for their AI models, and catalyze translational research in the field.

Acknowledgements

Key contributors to this project include Corbin Cunningham, Zaid Nabulsi, Ryan Najafi, Jie Yang, Charles Lau, Joseph R. Ledsam, Wenxing Ye, Diego Ardila, Scott M. McKinney, Rory Pilgrim, Hiroaki Saito, Yasuteru Shimamura, Mozziyar Etemadi, Yun Liu, David Melnick, Sunny Jansen, Nadia Harhen, David P. Nadich, Mikhail Fomitchev, Ziyad Helali, Shabir Adeel, Greg S. Corrado, Lily Peng, Daniel Tse, Shravya Shetty, Shruthi Prabhakara, Neeral Beladia, and Krish Eswaran. Thanks to Arnav Agharwal and Andrew Sellergren for their open sourcing support and Vivek Natarajan and Michael D. Howell for their feedback. Sincere appreciation also goes to the radiologists who enabled this work with their image interpretation and annotation efforts throughout the study, and Jonny Wong and Carli Sampson for coordinating the reader studies.