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 10100 publications
PaLI-X: On Scaling up a Multilingual Vision and Language Model
Josip Djolonga
Piotr Padlewski
Basil Mustafa
Carlos Riquelme
Sebastian Goodman
Yi Tay
Siamak Shakeri
Daniel Salz
Michael Tschannen
Mandar Joshi
Filip Pavetić
Gang Li
Anurag Arnab
Yuanzhong Xu
Keran Rong
Neil Houlsby
Computer Vision and Pattern Recognition Conference (CVPR) (2024)
Preview abstract
We explore the boundaries of scaling up a multilingual vision and language model, both in terms of size of the components and the breadth of its training task mixture. Our model achieves new levels of performance on a wide-range of varied and complex tasks, including multiple image-based captioning and question-answering tasks, image-based document understanding and few-shot (in-context) learning, as well as object detection, video question answering, and video captioning. Our model advances the state-of-the-art on most vision-and-language benchmarks considered (20+ of them). Finally, we observe emerging capabilities, such as complex counting and multilingual object detection, tasks that are not explicitly in the training mix.
View details
Factual and Personalized Recommendation Language Modeling with Reinforcement Learning
Jihwan Jeong
Mohammad Ghavamzadeh
Proceedings of the First Conference on Language Modeling (COLM-24), Philadelphia (2024)
Preview abstract
Recommender systems (RSs) play a central role in connecting users to products, content and services by matching candidate items to users based on their preferences. While existing RSs often rely on implicit user feedback on recommended items (e.g., clicks, watches, ratings), conversational recommender systems are interacting with users to provide tailored recommendations in natural language. In this work, we aim to develop a recommender language model (LM) that is capable of generating compelling endorsement presentations of relevant items to users, to better explain the details of the items, to connect the items with users’ preferences, and to enhance the likelihood of users accepting recommendations. Specifically, such an LLM-based recommender can understand users’ preferences from users’ RS embeddings summarizing feedback history, output corresponding responses that not only are factually-grounded, but also explain whether these items satisfy users’ preferences in a convincing manner. The pivotal question is how one can gauge the performance of such a LLM recommender. Equipped with a joint reward function that measures factual consistency, convincingness, and personalization, not only can we evaluate the efficacies of different recommender LMs, but we can also utilize this metric as a form of AI feedback to fine-tune our LLM agent via reinforcement learning (RL). Building upon the MovieLens movie recommendation benchmark, we developed a novel conversational recommender delivering personalized movie narratives to users. This work lays the groundwork for recommendation systems that prioritize individualized user experiences without compromising on transparency and integrity.
View details
Preview abstract
Machine learning has a pseudoscience problem. An abundance of ethical issues arising from the use of machine learning (ML)-based technologies—by now, well documented—is inextricably entwined with the systematic epistemic misuse of these tools. We take a recent resurgence of deep learning-assisted physiognomic research as a case study in the relationship between ML-based pseudoscience and attendant social harms—the standard purview of “AI ethics.” In practice, the epistemic and ethical dimensions of ML misuse often arise from shared underlying reasons and are resolvable by the same pathways. Recent use of ML toward the ends of predicting protected attributes from photographs highlights the need for philosophical, historical, and domain-specific perspectives of particular sciences in the prevention and remediation of misused ML.
View details
Preview abstract
The InterPlanetary File System (IPFS) is on its way to becoming the backbone of the next generation of the web. However, it suffers from several performance bottlenecks, particularly on the content retrieval path, which are often difficult to debug. This is because content retrieval involves multiple peers on the decentralized network and the issue could lie anywhere in the network. Traditional debugging tools are insufficient to help web developers who face the challenge of slow loading websites and detrimental user experience. This limits the adoption and future scalability of IPFS.
In this paper, we aim to gain valuable insights into how content retrieval requests propagate within the IPFS network as well as identify potential performance bottlenecks which could lead to opportunities for improvement. We propose a custom tracing framework that generates and manages traces for crucial events that take place on each peer during content retrieval. The framework leverages event semantics to build a timeline of each protocol involved in the retrieval, helping developers pinpoint problems. Additionally, it is resilient to malicious behaviors of the peers in the decentralized environment.
We have implemented this framework on top of an existing IPFS implementation written in Java called Nabu. Our evaluation shows that the framework can identify network delays and issues with each peer involved in content retrieval requests at a very low overhead.
View details
Augmentations vs Algorithms: What Works in Self-Supervised Learning
Warren Morningstar
Alex Bijamov
Chris Duvarney
Luke Friedman
Neha Kalibhat
Philip Mansfield
Renan Rojas-Gomez
Karan Singhal
Bradley Green
Sushant Prakash
Arxiv (2024) (to appear)
Preview abstract
We study the relative effects of data augmentations, pretraining algorithms, and model architectures in Self-Supervised Learning (SSL). While the recent literature in this space leaves the impression that the pretraining algorithm is of critical importance to performance, understanding its effect is complicated by the difficulty in making objective and direct comparisons between methods. We propose a new framework which unifies many seemingly disparate SSL methods into a single shared template. Using this framework, we identify aspects in which methods differ and observe that in addition to changing the pretraining algorithm, many works also use new data augmentations or more powerful model architectures. We compare several popular SSL methods using our framework and find that many algorithmic additions, such as prediction networks or new losses, have a minor impact on downstream task performance (often less than 1%), while enhanced augmentation techniques offer more significant performance improvements (2−4%). Our findings challenge the premise that SSL is being driven primarily by algorithmic improvements, and suggest instead a bitter lesson for SSL: that augmentation diversity and data / model scale are more critical contributors to recent advances in self-supervised learning.
View details
Preview abstract
Browser fingerprinting is often associated with cross-site user tracking, a practice that many browsers (e.g., Safari, Brave, Edge, Firefox and Chrome) want to block. However, less is publicly known about its uses to enhance online safety, where it can provide an additional security layer against service abuses (e.g., in combination with CAPTCHAs) or during user authentication. To the best of our knowledge, no fingerprinting defenses deployed thus far consider this important distinction when blocking fingerprinting attempts, so they might negatively affect website functionality and security.
To address this issue we make three main contributions. First, we propose and evaluate a novel machine learning-based method to automatically identify authentication pages (i.e. sign-in and sign-up pages). Our algorithm -- which relies on a hybrid unsupervised/supervised approach -- achieves 96-98% precision and recall on a large, manually-labelled dataset of 10,000 popular sites. Second, we compare our algorithm with other methods from prior works on the same dataset, showing that it significantly outperforms all of them (+83% F1-score). Third, we quantify the prevalence of fingerprinting scripts across sign-in and sign-up pages (9.2%) versus those executed on other pages (8.9%); while the rates of fingerprinting are similar, home pages and authentication pages differ in the third-party scripts they include and how often these scripts are labeled as tracking. We also highlight the substantial differences in fingerprinting behavior on login and sign-up pages.
Our work sheds light on the complicated reality that fingerprinting is used to both protect user security and invade user privacy, and that this dual nature must be considered by fingerprinting mitigations.
View details
Towards Generalist Biomedical AI
Danny Driess
Andrew Carroll
Chuck Lau
Ryutaro Tanno
Ira Ktena
Anil Palepu
Basil Mustafa
Aakanksha Chowdhery
Simon Kornblith
Philip Mansfield
Sushant Prakash
Renee Wong
Sunny Virmani
Sara Mahdavi
Bradley Green
Ewa Dominowska
Joelle Barral
Karan Singhal
Pete Florence
NEJM AI (2024)
Preview abstract
BACKGROUND: Medicine is inherently multimodal, requiring the simultaneous interpretation and integration of insights between many data modalities spanning text, imaging, genomics, and more. Generalist biomedical artificial intelligence systems that flexibly encode, integrate, and interpret these data might better enable impactful applications ranging from scientific discovery to care delivery.
METHODS: To catalyze development of these models, we curated MultiMedBench, a new multimodal biomedical benchmark. MultiMedBench encompasses 14 diverse tasks, such as medical question answering, mammography and dermatology image interpretation, radiology report generation and summarization, and genomic variant calling. We then introduced Med-PaLM Multimodal (Med-PaLM M), our proof of concept for a generalist biomedical AI system that flexibly encodes and interprets biomedical data including clinical language, imaging, and genomics with the same set of model weights. To further probe the capabilities and limitations of Med-PaLM M, we conducted a radiologist evaluation of model-generated (and human) chest x-ray reports.
RESULTS: We observed encouraging performance across model scales. Med-PaLM M reached performance competitive with or exceeding the state of the art on all MultiMedBench tasks, often surpassing specialist models by a wide margin. In a side-by-side ranking on 246 retrospective chest x-rays, clinicians expressed a pairwise preference for Med-PaLM Multimodal reports over those produced by radiologists in up to 40.50% of cases, suggesting potential clinical utility.
CONCLUSIONS: Although considerable work is needed to validate these models in real-world cases and understand if cross-modality generalization is possible, our results represent a milestone toward the development of generalist biomedical artificial intelligence systems.
View details
Bridging the Preference Gap between Retrievers and LLMs
Zixuan Ke
Qiaozhu Mei
Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (2024) (to appear)
Preview abstract
Large Language Models (LLMs) have demonstrated superior results across a wide range of tasks, and Retrieval-augmented Generation (RAG) is an effective way to enhance the performance by locating relevant information and placing it into the context window of the LLM. However, the relationship between retrievers and LLM in a RAG is still under-investigated. Most existing work treats the retriever and the LLM as independent components and leaves a gap between retrieving human-"friendly" information and assembling a LLM-"friendly" context. In this work, we examine a novel bridge mechanism. We validate the ranking and selection assumptions of retrievers in the context of RAG and propose a framework that chains together supervised and reinforcement learning to train a bridge model that optimizes the connection between the retriever and the LLM. Empirical results demonstrate the effectiveness of our method in both question-answering and personalized generation tasks.
View details
Preview abstract
Neural embedding models have become a fundamental component of modern information retrieval (IR) pipelines. These models produce a single embedding x ∈ R^d per data-point, allowing for fast retrieval via highly optimized maximum inner product search (MIPS) algorithms. Recently, beginning with the landmark ColBERT paper, multi-vector models, which produce a set of embedding per data point, have achieved markedly superior performance for IR tasks. Unfortunately, using these models for IR is computationally expensive due to the increased complexity of multi-vector retrieval and scoring.
In this paper, we introduce MUVERA (Multi-Vector Retrieval Algorithm), a retrieval mechanism which reduces multi-vector similarity search to single-vector similarity search. This enables the usage of off-the-shelf MIPS solvers for multi-vector retrieval. MUVERA asymmetrically generates Fixed Dimensional Encodings (FDEs) of queries and documents, which are vectors whose inner product approximates multi-vector similarity. We prove that FDEs give high-quality ε-approximations, thus providing the first single-vector proxy for multi-vector similarity with theoretical guarantees. Empirically, we find that FDEs achieve the same recall as prior state-of-the-art heuristics while retrieving 2-5× fewer candidates. Compared to prior state of the art implementations, MUVERA achieves consistently good end-to-end recall and latency across a diverse set of the BEIR retrieval datasets, achieving an average of 10% improved recall with 90% lower latency.
View details
AI-assisted Assessment of Coding Practices in Industrial Code Review
Ivan Budiselic
Malgorzata (Gosia) Salawa
Juanjo Carin
Jovan Andonov
Mateusz Lewko
Rene Just
Preview abstract
Modern code review is a process in which incremental code contributions made by one software developer are reviewed by one or more peers before it is committed to the version control system. An important element of modern code review is verifying that the code under review adheres to style guidelines and best practices of the corresponding programming language. Some of these rules are universal and can be checked automatically or enforced via code formatters. Other rules, however, are context-dependent and the corresponding checks are commonly left to developers who are experts in the given programming language and whose time is expensive. Many automated systems have been developed that attempt to detect various rule violations without any human intervention. Historically, such systems implement targeted analyses and were themselves expensive to develop. This paper presents AutoCommenter, a system that uses a state of the art large language model to automatically learn and enforce programming language best practices. We implemented AutoCommenter for four programming languages: C++, Java, Python and Go. We evaluated its performance and adoption in a large industrial setting. Our evaluation shows that a model that automatically learns language best practices is feasible and has a measurable positive impact on the developer workflow. Additionally, we present the challenges we faced when deploying such a model to tens of thousands of developers and provide lessons we learned for any practitioners that would like to replicate the work or build on top of it.
View details
A Chain-of-Thought Is as Strong as Its Weakest Link: A Benchmark for Verifiers of Reasoning Chains
Alon Jacovi
Or Honovich
Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (2024), pp. 4615–4634
Preview abstract
Prompting language models to provide step-by-step answers (e.g., “Chain-of-Thought”) is the prominent approach for complex reasoning tasks, where more accurate reasoning chains typically improve downstream task performance. Recent literature discusses automatic methods to verify reasoning to evaluate and improve their correctness. However, no fine-grained step-level datasets are available to enable thorough evaluation of such verification methods, hindering progress in this direction. We introduce REVEAL: Reasoning Verification Evaluation, a dataset to benchmark automatic verifiers of complex Chain-of-Thought reasoning in open-domain question-answering settings. REVEAL includes comprehensive labels for the relevance, attribution to evidence passages, and logical correctness of each reasoning step in a language model’s answer, across a variety of datasets and state-of-the-art language models. Evaluation on REVEAL shows that verifiers struggle at verifying reasoning chains — in particular, verifying logical correctness and detecting contradictions. Available at https://reveal-dataset.github.io/.
View details
Connecting Language Technologies with Rich, Diverse Data Sources Covering Thousands of Languages
Sebastian Ruder
Julia Kreutzer
Clara Rivera
Ishank Saxena
Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)
Preview abstract
Contrary to common belief, there are rich and diverse data sources available for many thousands of languages, which can be used to develop technologies for these languages. In this paper, we provide an overview of some of the major online data sources, the types of data that they provide access to, potential applications of this data, and the number of languages that they cover. Even this covers only a small fraction of the data that exists; for example, printed books are published in many languages but few online aggregators exist.
View details
Preview abstract
Japanese text-to-pronunciation modelling is a notoriously data-intensive problem. Japanese data sources are often only partially annotated, and use different annotation standards for pronunciation and word segmentation. This talk introduces a set of techniques that enable ingesting data that may be partially annotated, use arbitrary word segmentations, and use a variety of pronunciation annotation standards.
View details
AI-Enhanced API Design: A New Paradigm in Usability and Efficiency
Mak Ahmad
David R Karger
Kwan-Liu Ma
CHI EA '24: Extended Abstracts of the 2024 CHI Conference on Human Factors in Computing Systems (2024)
Preview abstract
This study uses mixed methods to evaluate API design methods, focusing on design and consumption phases. Our goal was to understand the impact of API governance approaches on productivity and usability. A controlled developer experiment (n=34) demonstrated
a 10% increased requirement fulfillment using API Improvement Proposals (AIPs) and linter versus no protocols. Meanwhile, 73% of 33 surveyed API consumers preferred AIP-aligned designs for enhanced usability and comprehensibility. Complementing this, a
custom large language model called the API Architect received average expert ratings of just 5/10 for specification quality, revealing gaps versus manual design. The quantitative performance metrics combined with qualitative user feedback provide evidence from
multiple angles that strategically integrating industry best practices with maturing AI capabilities can meaningfully improve API design outcomes. This research offers empirical insights from developer and consumer perspectives to advance scholarly discourse
and industry practice regarding optimal API design workflows.
View details
Preview abstract
Google Cloud SQL customers encounter PostgreSQL bugs corrupting databases, rarely but reproducibly. This talk will cover use of tools, especially amcheck, to grasp these bugs sufficiently to write fixes and test cases. Those fixes are now part of core PostgreSQL. It will include lessons for avoiding such bugs in future PostgreSQL development. Finally, it will share a diagnostic feature wish list.
View details