Publications
Our teams aspire to make discoveries that impact everyone, and core to our approach is sharing our research and tools to fuel progress in the field.
Our teams aspire to make discoveries that impact everyone, and core to our approach is sharing our research and tools to fuel progress in the field.
Sort By
1 - 15 of 10292 publications
Validation of a Deep Learning Model for Diabetic Retinopathy on Patients with Young-Onset Diabetes
Tony Tan-Torres
Pradeep Praveen
Divleen Jeji
Arthur Brant
Xiang Yin
Lu Yang
Tayyeba Ali
Ilana Traynis
Dushyantsinh Jadeja
Rajroshan Sawhney
Sunny Virmani
Pradeep Venkatesh
Nikhil Tandon
Ophthalmology and Therapy (2025)
Preview abstract
Introduction
While many deep learning systems (DLSs) for diabetic retinopathy (DR) have been developed and validated on cohorts with an average age of 50s or older, fewer studies have examined younger individuals. This study aimed to understand DLS performance for younger individuals, who tend to display anatomic differences, such as prominent retinal sheen. This sheen can be mistaken for exudates or cotton wool spots, and potentially confound DLSs.
Methods
This was a prospective cross-sectional cohort study in a “Diabetes of young” clinic in India, enrolling 321 individuals between ages 18 and 45 (98.8% with type 1 diabetes). Participants had fundus photographs taken and the photos were adjudicated by experienced graders to obtain reference DR grades. We defined a younger cohort (age 18–25) and an older cohort (age 26–45) and examined differences in DLS performance between the two cohorts. The main outcome measures were sensitivity and specificity for DR.
Results
Eye-level sensitivity for moderate-or-worse DR was 97.6% [95% confidence interval (CI) 91.2, 98.2] for the younger cohort and 94.0% [88.8, 98.1] for the older cohort (p = 0.418 for difference). The specificity for moderate-or-worse DR significantly differed between the younger and older cohorts, 97.9% [95.9, 99.3] and 92.1% [87.6, 96.0], respectively (p = 0.008). Similar trends were observed for diabetic macular edema (DME); sensitivity was 79.0% [57.9, 93.6] for the younger cohort and 77.5% [60.8, 90.6] for the older cohort (p = 0.893), whereas specificity was 97.0% [94.5, 99.0] and 92.0% [88.2, 95.5] (p = 0.018). Retinal sheen presence (94% of images) was associated with DME presence (p < 0.0001). Image review suggested that sheen presence confounded reference DME status, increasing noise in the labels and depressing measured sensitivity. The gradability rate for both DR and DME was near-perfect (99% for both).
Conclusion
DLS-based DR screening performed well in younger individuals aged 18–25, with comparable sensitivity and higher specificity compared to individuals aged 26–45. Sheen presence in this cohort made identification of DME difficult for graders and depressed measured DLS sensitivity; additional studies incorporating optical coherence tomography may improve accuracy of measuring DLS DME sensitivity.
View details
Sufficient Context: A New Lens on Retrieval Augmented Generation Systems
Hailey Joren
Jianyi Zhang
Chun-Sung Ferng
Ankur Taly
International Conference on Learning Representations (ICLR) (2025)
Preview abstract
Augmenting LLMs with context leads to improved performance across many applications. Despite much research on Retrieval Augmented Generation (RAG) systems, an open question is whether errors arise because LLMs fail to utilize the context from retrieval or the context itself is insufficient to answer the query. To shed light on this, we develop a new notion of sufficient context, along with a method to classify instances that have enough information to answer the query. We then use sufficient context to analyze several models and datasets. By stratifying errors based on context sufficiency, we find that larger models with higher baseline performance (Gemini 1.5 Pro, GPT 4o, Claude 3.5) excel at answering queries when the context is sufficient, but often output incorrect answers instead of abstaining when the context is not. On the other hand, smaller models with lower baseline performance (Llama 3.1, Mistral 3, Gemma 2) hallucinate or abstain often, even with sufficient context. We further categorize cases when the context is useful, and improves accuracy, even though it does not fully answer the query and the model errs without the context. Building on our findings, we explore ways to reduce hallucinations in RAG systems, including a new selective generation method that leverages sufficient context information for guided abstention. Our method improves the fraction of correct answers among times where the model responds by 2--10% for Gemini, GPT, and Gemma.
View details
Preview abstract
We study the existence of almost fair and near-optimal solutions to a routing problem as defined in the seminal work of Rosenthal. We focus on the setting where multiple alternative routes are available for each potential request (which corresponds to a potential user of the network). This model captures a collection of diverse applications such as packet routing in communication networks, routing in road networks with multiple alternative routes, and the economics of transportation of goods.
Our recommended routes have provable guarantees in terms of both the total cost and fairness concepts such as approximate envy-freeness. We employ and appropriately combine tools from algorithmic game theory and fair division. Our results apply on two distinct models: the splittable case where the request is split among the selected paths (e.g., routing a fleet of trucks) and the unsplittable case where the request is assigned to one of its designated paths (e.g., a single user request). Finally, we conduct an empirical analysis to test the performance of our approach against simpler baselines using the real world road network of New York City.
View details
Mufu: Multilingual Fused Learning for Low- Resource Translation with LLM
Zheng Lim
Honglin Yu
Trevor Cohn
International Conference on Learning Representations (ICLR) 2025
Preview abstract
Multilingual large language models (LLMs) are great translators, but this is largely limited to high-resource languages. For many LLMs, translating in and out of low-resource languages remains a challenging task. To maximize data efficiency in this low-resource setting, we introduce Mufu, which includes a selection of automatically generated multilingual candidates and an instruction to correct inaccurate translations in the prompt. Mufu prompts turn a translation task into a postediting one, and seek to harness the LLM's reasoning capability with auxiliary translation candidates, from which the model is required to assess the input quality, align the semantics cross-lingually, copy from relevant inputs and override instances that are incorrect. Our experiments on En-XX translations over the Flores-200 dataset show LLMs finetuned against Mufu-style prompts are robust to poor quality auxiliary translation candidates, achieving performance superior to NLLB 1.3B distilled model in 64% of low- and very-low-resource language pairs. We then distill these models to reduce inference cost, while maintaining on average 3.1 chrF improvement over finetune-only baseline in low-resource translations.
View details
Triaging mammography with artificial intelligence: an implementation study
Sarah M. Friedewald
Sunny Jansen
Fereshteh Mahvar
Timo Kohlberger
David V. Schacht
Sonya Bhole
Dipti Gupta
Scott Mayer McKinney
Stacey Caron
David Melnick
Mozziyar Etemadi
Samantha Winter
Alejandra Maciel
Luca Speroni
Martha Sevenich
Arnav Agharwal
Rubin Zhang
Gavin Duggan
Shiro Kadowaki
Atilla Kiraly
Jie Yang
Basil Mustafa
Krish Eswaran
Shravya Shetty
Breast Cancer Research and Treatment (2025)
Preview abstract
Purpose
Many breast centers are unable to provide immediate results at the time of screening mammography which results in delayed patient care. Implementing artificial intelligence (AI) could identify patients who may have breast cancer and accelerate the time to diagnostic imaging and biopsy diagnosis.
Methods
In this prospective randomized, unblinded, controlled implementation study we enrolled 1000 screening participants between March 2021 and May 2022. The experimental group used an AI system to prioritize a subset of cases for same-visit radiologist evaluation, and same-visit diagnostic workup if necessary. The control group followed the standard of care. The primary operational endpoints were time to additional imaging (TA) and time to biopsy diagnosis (TB).
Results
The final cohort included 463 experimental and 392 control participants. The one-sided Mann-Whitney U test was employed for analysis of TA and TB. In the control group, the TA was 25.6 days [95% CI 22.0–29.9] and TB was 55.9 days [95% CI 45.5–69.6]. In comparison, the experimental group's mean TA was reduced by 25% (6.4 fewer days [one-sided 95% CI > 0.3], p<0.001) and mean TB was reduced by 30% (16.8 fewer days; 95% CI > 5.1], p=0.003). The time reduction was more pronounced for AI-prioritized participants in the experimental group. All participants eventually diagnosed with breast cancer were prioritized by the AI.
Conclusions
Implementing AI prioritization can accelerate care timelines for patients requiring additional workup, while maintaining the efficiency of delayed interpretation for most participants. Reducing diagnostic delays could contribute to improved patient adherence, decreased anxiety and addressing disparities in access to timely care.
View details
Towards Conversational AI for Disease Management
Anil Palepu
Khaled Saab
David Stutz
Kavita Kulkarni
Sara Mahdavi
Joelle Barral
James Manyika
Ryutaro Tanno
Adam Rodman
arXiv (2025)
Preview abstract
While large language models (LLMs) have shown promise in diagnostic dialogue, their capabilities for effective management reasoning - including disease progression, therapeutic response, and safe medication prescription - remain under-explored. We advance the previously demonstrated diagnostic capabilities of the Articulate Medical Intelligence Explorer (AMIE) through a new LLM-based agentic system optimised for clinical management and dialogue, incorporating reasoning over the evolution of disease and multiple patient visit encounters, response to therapy, and professional competence in medication prescription. To ground its reasoning in authoritative clinical knowledge, AMIE leverages Gemini's long-context capabilities, combining in-context retrieval with structured reasoning to align its output with relevant and up-to-date clinical practice guidelines and drug formularies. In a randomized, blinded virtual Objective Structured Clinical Examination (OSCE) study, AMIE was compared to 21 primary care physicians (PCPs) across 100 multi-visit case scenarios designed to reflect UK NICE Guidance and BMJ Best Practice guidelines. AMIE was non-inferior to PCPs in management reasoning as assessed by specialist physicians and scored better in both preciseness of treatments and investigations, and in its alignment with and grounding of management plans in clinical guidelines. To benchmark medication reasoning, we developed RxQA, a multiple-choice question benchmark derived from two national drug formularies (US, UK) and validated by board-certified pharmacists. While AMIE and PCPs both benefited from the ability to access external drug information, AMIE outperformed PCPs on higher difficulty questions. While further research would be needed before real-world translation, AMIE's strong performance across evaluations marks a significant step towards conversational AI as a tool in disease management.
View details
Gemini & Physical World: Large Language Models Can Estimate the Intensity of Earthquake Shaking from Multi-Modal Social Media Posts
Marc Stogaitis
Tajinder Gadh
Richard Allen
Alexei Barski
Robert Bosch
Patrick Robertson
Youngmin Cho
Nivetha Thiruverahan
Aman Raj
Geophysical Journal International (2025), ggae436
Preview abstract
This paper presents a novel approach for estimating the ground shaking intensity using real-time social media data and CCTV footage. Employing the Gemini 1.5 Pro’s (Reid et al. 2024) model, a multi-modal language model, we demonstrate the ability to extract relevant information from unstructured data utilizing generative AI and natural language processing. The model’s output, in the form of Modified Mercalli Intensity (MMI) values, align well with independent observational data. Furthermore, our results suggest that beyond its advanced visual and auditory understanding abilities, Gemini appears to utilize additional sources of knowledge, including a simplified understanding of the general relationship between earthquake magnitude, distance, and MMI intensity, which it presumably acquired during its training, in its reasoning and decision-making processes. These findings raise intriguing questions about the extent of Gemini's general understanding of the physical world and its phenomena. Gemini’s ability to generate results consistent with established scientific knowledge highlights the potential of LLMs like Gemini in augmenting our understanding of complex physical phenomena such as earthquakes. More specifically, the results of this study highlight the potential of LLMs like Gemini to revolutionize citizen seismology by enabling rapid, effective, and flexible analysis of crowdsourced data from eyewitness accounts for assessing earthquake impact and providing crisis situational awareness. This approach holds a great promise for improving early warning systems, disaster response, and overall resilience in earthquake-prone regions. This study provides a significant step toward harnessing the power of social media and AI for earthquake disaster mitigation.
View details
Preview abstract
We present a scalable and agile approach for ads image content moderation at Google, addressing the challenges of moderating massive volumes of ads with diverse content and evolving policies. The proposed method utilizes human-curated textual descriptions and cross-modal text-image co-embeddings to enable zero-shot classification of policy violating ads images, bypassing the need for extensive supervised training data and human labeling. By leveraging large language models (LLMs) and user expertise, the system generates and refines a comprehensive set of textual descriptions representing policy guidelines. During inference, co-embedding similarity between incoming images and the textual descriptions serves as a reliable signal for policy violation detection, enabling efficient and adaptable ads content moderation. Evaluation results demonstrate the efficacy of this framework in significantly boosting the detection of policy violating content.
View details
Preview abstract
Augmenting LLMs with context leads to improved performance across many applications. Despite much research on Retrieval Augmented Generation (RAG) systems, an open question is whether errors arise because LLMs fail to utilize the context from retrieval or the context itself is insufficient to answer the query. To shed light on this, we develop a new notion of sufficient context, along with a way to classify instances that have enough information to answer the query. We then use sufficient context to analyze several models and datasets. By stratifying errors based on context sufficiency, we find that proprietary LLMs (Gemini, GPT, Claude) excel at answering queries when the context is sufficient, but often output incorrect answers instead of abstaining when the context is not. On the other hand, open-source LLMs (Llama, Mistral, Gemma) hallucinate or abstain often, even with sufficient context. We further categorize cases when the context is useful, and improves accuracy, even though it does not fully answer the query and the model errs without the context. Building on our findings, we explore ways to reduce hallucinations in RAG systems, including a new selective generation method that leverages sufficient context information for guided abstention. Our method improves the fraction of correct answers among times where the model responds by 2--10% for Gemini, GPT, and Gemma.
View details
Preview abstract
Despite the advent of legislation such as the General Data Protection Regulation (GDPR) with its associated "Right to be Forgotten" (RTBF), few, if any, studies have measured user reactions to realistic edge cases with public-interest content. Surveying both users covered by and excluded from RTBF, this vignette-based survey experiment sought to better understand how users think of delisting content from search engine results and what factors influence user perceptions. While leaving information accessible in search engine results generally leads to warmer feelings towards those search engines than delisting it, we find that users do prefer different outcomes
depending on contextual elements specific to given cases. We also find that whether a country has active RTBF legislation does seem to be associated with both knowledge and attitudes about RTBF, but is unlikely to explain all of it. These results indicate a complex context around removing public-interest content from search engines’ results; it is essential that experts sensitive to local context perform the review in order to ensure that removal requests are handled in a way that meets users’ expectations.
View details
Beyond Touchscreens: Dynamic and Multimodal Interaction Needs
Melissa Barnhart Wantland
Mai Kobori
Universal Access in Human-Computer Interaction, Springer-Verlag (2025) (to appear)
Preview abstract
Today’s smartphone interactions are typically designed with one primary preset, accompanied by customization settings that can be manually adjusted. To promote the creation of contextually aware experiences, researchers have highlighted the factors that influence mobile device usage in the ability-based design framework. This paper expands upon existing frameworks and contributes to an empirical understanding of smartphone accessibility. Through a 10-day longitudinal diary study and video interview with 24 individuals who do and do not identify as having a disability, the research also illustrates the reactions of reattempt, adaptation, and avoidance, which were used in response to a lack of smartphone accessibility. Despite experiencing scenarios where accessibility settings could be leveraged, 20 out of 24 participants did not use accessibility settings on their smartphone. A total of 12 out of 24 participants tried accessibility settings on their smartphones, however identifying accessibility was not for them. This work highlights the need to shift current design practices to better serve the accessibility community.
View details
SSDTrain: Faster Large Language Model Training Using SSD-Based Activation Offloading
Kun Wu
Jeongmin Brian Park
Mert Hidayetoğlu
Vikram Sharma Mailthody
Sitao Huang
Steven Lumetta
Wen-mei Hwu
Design Automation Conference (DAC) (2025)
Preview abstract
The scaling up of Large Language Models (LLMs) demands more memory than current GPUs can provide, hindering the training process. To address this challenge, we propose SSDTrain to efficiently offload activations, the intermediate tensors produced during LLM training, to SSDs. This approach reduces GPU memory usage without impacting performance by adaptively overlapping data transfers with computation. SSDTrain is compatible with popular deep learning frameworks like PyTorch, Megatron, and DeepSpeed, and it employs techniques such as tensor deduplication, forwarding, and adaptive offloading to further enhance efficiency. We conduct extensive experiments on Llama, BERT, and T5. Results demonstrate that SSDTrain effectively reduces 45% of the activation peak memory usage. It can perfectly overlap the IO with the computation without introducing performance penalty. SSDTrain can achieve a performance boost of up to 31% compared to the conventional training strategy using the same GPU systems.
View details
Preview abstract
The problem of contract design addresses the challenge of moral hazard in principle-agent setups. The agent exerts costly efforts that produce a random outcome with an associated reward for the principal. Moral hazard refers to the tension that the principal cannot observe the agent’s effort level hence needs to incentivize the agent only through rewarding the realized effort outcome, i.e., the contract. Bayesian contract design studies the principal’s design problem of an optimal contract when facing an unknown agent characterized by a private Bayesian type. In its most general form, the agent’s type is inherently “multi-parameter” and can arbitrarily affect both the agent’s productivity and effort costs. In contrast, a natural single-parameter setting of much recent interest simplifies the agent’s type to a single value that describes the agent’s cost per unit of effort, whereas agents’ efforts are assumed to be equally
productive.
The main result of this paper is an almost approximation-preserving polynomial-time reduction from the most general multi-parameter Bayesian contract design (BCD) to single-parameter BCD. That is, for any multi-parameter BCD instance I^M, we construct a single-parameter instance I^S such that any β-approximate contract (resp. menu of contracts) of I^S can in turn be converted to a (β − ϵ)-approximate contract (resp. menu of contracts) of I^M. The reduction is in time polynomial in the input size and log(1/ϵ); moreover, when β = 1 (i.e., the given single-parameter solution is exactly optimal), the dependence on 1/ϵ can be removed, leading to a polynomial-time exact reduction. This efficient reduction is somewhat surprising because in the closely related problem of Bayesian mechanism design, a polynomial-time reduction from multi-parameter to single-parameter setting is believed to not exist. Our result demonstrates the intrinsic difficulty of addressing moral hazard in Bayesian contract design, regardless of being single-parameter or multi-parameter.
As byproducts, our reduction answers two open questions in recent literature of algorithmic contract design: (a) it implies that optimal contract design in single-parameter BCD is not in APX unless P=NP even when the agent’s type distribution is regular, answering the open question of [3] in the negative; (b) it implies that the principal’s (order-wise) tight utility gap between using a menu of contracts and a single contract is Θ(n) where n is the number of actions, answering the major open question of [27] for the single-parameter case.
View details
H2E: Hand, Head, Eye: A Multimodal Cascade of Natural Inputs
Khushman Patel
Ken Pfeuffer
Hans Gellersen
IEEE VR (2025)
Preview abstract
Eye-based interaction techniques for extended reality, such as gaze and pinch, are simple to use however suffer from input precision issues. We present H2E, a fine and coarse-grained pointing technique that cascades Hand, Head, and Eye inputs. As users initiate a pinch gesture, a cursor appears at the gaze point that can be dragged by head pointing before pinch confirmation. This has the potential advantage that it can add a precision component without changing the semantics of the technique. In this paper, we describe the design and implementation of the technique. Furthermore, we present an evaluation of our method in a Fitts-based user study, exploring the speed-accuracy trade-offs against a gaze and pinch interaction baseline.
View details
Preview abstract
Mainstream artificial neural network models, such as Deep Neural Networks (DNNs) are computation-heavy and energy-hungry. Weightless Neural Networks (WNNs) are natively built with RAM-based neurons and represent an entirely distinct type of neural network computing compared to DNNs. WNNs are extremely low-latency, low-energy, and suitable for efficient, accurate, edge inference. The WNN approach derives an implicit inspiration from the decoding process observed in the dendritic trees of biological neurons, making neurons based on Random Access Memories (RAMs) and/or Lookup Tables (LUTs) ready-to-deploy neuromorphic digital circuits. Since FPGAs are abundant in LUTs, LUT based WNNs are a natural fit for implementing edge inference in FPGAs.
WNNs has been demonstrated to be an energetically efficient AI model, both in software, as well as in hardware. For instance, the most recent DWN – Differential Weightless Neural Network – model demonstrates up to 135× reduction in energy costs in FPGA implementations compared to other multiplication-free approaches, such as binary neural networks (BNNs) and DiffLogicNet, up to 9% higher accuracy in deployments on constrained devices, and culminate in up to 42.8× reduction in circuit area for ultra-low-cost chip implementations. This tutorial will help participants understand how WNNs work, why WNNs were underdogs for such a long time, and be introduced to the most recent members of the WNN family, such as BTHOWeN , LogicWiSARD, COIN, ULEEN and DWN, and contrast to BNNs and LogicNets.
View details
