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 10128 publications
Preview abstract
Detecting offensive content in text is an increasingly central challenge for both social-media platforms and AI-driven technologies. However offensiveness remains a subjective phenomenon as perspectives differ across sociodemographic characteristics, as well as cultural norms and moral values. This intricacy is largely ignored in the current AI-focused approaches for detecting offensiveness or related concepts such as hate speech and toxicity detection. We frame the task of determining offensiveness as essentially a matter of moral judgment --- deciding the boundaries of ethically wrong vs. right language to be used or generated within an implied set of sociocultural norms. In this paper, we investigate how judgment of offensiveness varies across diverse global cultural regions, and the crucial role of moral values in shaping these variations. Our findings highlight substantial cross-cultural differences in perceiving offensiveness, with moral concerns about Caring and Purity as the mediating factor driving these differences. These insights are of importance as AI safety protocols, shaped by human annotators' inputs and perspectives, embed their moral values which do not align with the notions of right and wrong in all contexts, and for all individuals.
View details
Efficient data generation for source-grounded information-seeking dialogs: A use case for meeting transcripts
Lotem Golany
Maya Mamo
Nimrod Parasol
Omer Vandsburger
Nadav Bar
Ido Dagan
Findings of the Association for Computational Linguistics: EMNLP 2024, Association for Computational Linguistics, Miami, Florida, USA, pp. 1908-1925
Preview abstract
Automating data generation with Large Language Models (LLMs) has become increasingly popular. In this work, we investigate the feasibility and effectiveness of LLM-based data generation in the challenging setting of source-grounded information-seeking dialogs, with response attribution, over long documents. Our source texts consist of long and noisy meeting transcripts, adding to the task complexity. Since automating attribution remains difficult, we propose a semi-automatic approach: dialog queries and responses are generated with LLMs, followed by human verification and identification of attribution spans. Using this approach, we created MISeD – Meeting Information Seeking Dialogs dataset – a dataset of information-seeking dialogs focused on meeting transcripts. Models finetuned with MISeD demonstrate superior performance compared to off-the-shelf models, even those of larger size. Finetuning on MISeD gives comparable response generation quality to finetuning on fully manual data, while improving attribution quality and reducing time and effort.
View details
Preview abstract
We present Prequal (\emph{Probing to Reduce Queuing and Latency}), a load balancer
for distributed multi-tenant systems. Prequal aims to minimize
real-time request latency in the presence of heterogeneous server
capacities and non-uniform, time-varying antagonist load. It actively probes
server load to leverage the \emph{power of $d$ choices}
paradigm, extending it with asynchronous and reusable probes. Cutting
against received wisdom, Prequal does not balance CPU load, but instead
selects servers according to estimated latency and active requests-in-flight
(RIF). We explore its major design features on a testbed system
and evaluate it on YouTube, where it has been deployed for more than two years. Prequal has dramatically decreased tail latency, error rates, and resource use, enabling YouTube and
other production systems at Google to run at much higher utilization.
View details
Preview abstract
Learned reweighting (LRW) approaches to supervised learning use an optimization criterion to assign weights for training instances, in order to maximize performance on a representative validation dataset. We pose and formalize the problem of optimized selection of the validation set used in LRW training, to improve classifier generalization. In particular, we show that using hard-to-classify instances in the validation set has both a theoretical connection to, and strong empirical evidence of generalization. We provide an efficient algorithm for training this meta-optimized model, as well as a simple train-twice heuristic for careful comparative study. We demonstrate that LRW with easy validation data performs consistently worse than LRW with hard validation data, establishing the validity of our meta-optimization problem. Our proposed algorithm outperforms a wide range of baselines on a range of datasets and domain shift challenges (Imagenet-1K, CIFAR-100, Clothing-1M, CAMELYON, WILDS, etc.), with ~1% gains using VIT-B on Imagenet. We also show that using naturally hard examples for validation (Imagenet-R / Imagenet-A) in LRW training for Imagenet improves performance on both clean and naturally hard test instances by 1-2%. Secondary analyses show that using hard validation data in an LRW framework improves margins on test data, hinting at the mechanism underlying our empirical gains. We believe this work opens up new research directions for the meta-optimization of meta-learning in a supervised learning context.
View details
Translatotron 3: Speech to Speech Translation with Monolingual Data
Alon Levkovitch
Yifan Ding
Chulayuth Asawaroengchai
Preview abstract
This paper presents a novel approach to train a direct speech-to-speech translation model from monolingual datasets only in a fully unsupervised manner. The proposed approach combines back-translation, denoising autoencoder, and unsupervised embedding mapping techniques to achieve this goal. We demonstrate the effectiveness of the proposed approach by comparing it against a cascaded baseline using two Spanish and English datasets. The proposed approach achieved a significant improvement over the cascaded baseline on synthesized unpaired conversational and synthesized Common Voice $11$ datasets.
View details
A Decade of Privacy-Relevant Android App Reviews: Large Scale Trends
Omer Akgul
Michelle Mazurek
Benoit Seguin
Preview abstract
We present an analysis of 12 million instances of privacy-relevant reviews publicly visible on the Google Play Store that span a 10 year period. By leveraging state of the art NLP techniques, we examine what users have been writing about privacy along multiple dimensions: time, countries, app types, diverse privacy topics, and even across a spectrum of emotions. We find consistent growth of privacy-relevant reviews, and explore topics that are trending (such as Data Deletion and Data Theft), as well as those on the decline (such as privacy-relevant reviews on sensitive permissions). We find that although privacy reviews come from more than 200 countries, 33 countries provide 90% of privacy reviews. We conduct a comparison across countries by examining the distribution of privacy topics a country’s users write about, and find that geographic proximity is not a reliable indicator that nearby countries have similar privacy perspectives. We uncover some countries with unique patterns and explore those herein. Surprisingly, we uncover that it is not uncommon for reviews that discuss privacy to be positive (32%); many users express pleasure about privacy features within apps or privacy-focused apps. We also uncover some unexpected behaviors, such as the use of reviews to deliver privacy disclaimers to developers. Finally, we demonstrate the value of analyzing app reviews with our approach as a complement to existing methods for understanding users' perspectives about privacy.
View details
Developer Ecosystems for Software Safety
Commun. ACM, 67 (2024), 52–60
Preview abstract
This paper reflects on work at Google over the past decade to address common types of software safety and security defects. Our experience has shown that software safety is an emergent property of the software and tooling ecosystem it is developed in and the production environment into which it is deployed. Thus, to effectively prevent common weaknesses at scale, we need to shift-left the responsibility for ensuring safety and security invariants to the end-to-end developer ecosystem, that is, programming languages, software libraries, application frameworks, build and deployment tooling, the production platform and its configuration surfaces, and so forth.
Doing so is practical and cost effective when developer ecosystems are designed with application archetypes in mind, such as web or mobile apps: The design of the developer ecosystem can address threat model aspects that apply commonly to all applications of the respective archetype, and investments to ensure safety invariants at the ecosystem level amortize across many applications.
Applying secure-by-design principles to developer ecosystems at Google has achieved drastic reduction and in some cases near-zero residual rates of common classes of defects, across hundreds of applications being developed by thousands of developers.
View details
Preview abstract
Most of our interactions with digital content currently occur inside 2D screens, however moving from that format to immersive setups brings a paradigm shift. From content inside the screen to users inside the content. This change requires a revisit to how we blend the analog and the digital and how we transfer content between the two modes. Perhaps it even asks for new guidelines too. While different solutions appear in the space, the dynamic range only seems to widen. We can start to see what works and what does not work so well, in an empirical or ethnographic approach, beyond laboratory studies. But if we want to accelerate adoption we need to further the understanding on how current tasks can be improved. How this new form of interaction can increase their productivity. In this paper we focus on analyzing and converging what we think works, and envisioning how this new set of immersive devices and interactions can enable productivity beyond already existing tools.
View details
First Passage Percolation with Queried Hints
Kritkorn Karntikoon
Aaron Schild
Yiheng Shen
Ali Sinop
AISTATS (2024)
Preview abstract
Optimization problems are ubiquitous throughout the modern world. In many of these applications, the input is inherently noisy and it is expensive to probe all of the noise in the input before solving the relevant optimization problem. In this work, we study how much of that noise needs to be queried in order to obtain an approximately optimal solution to the relevant problem. We focus on the shortest path problem in graphs, where one may think of the noise as coming from real-time traffic. We consider the following model: start with a weighted base graph $G$ and multiply each edge weight by an independently chosen, uniformly random number in $[1,2]$ to obtain a random graph $G'$. This model is called \emph{first passage percolation}. Mathematicians have studied this model extensively when $G$ is a $d$-dimensional grid graph, but the behavior of shortest paths in this model is still poorly understood in general graphs. We make progress in this direction for a class of graphs that resembles real-world road networks. Specifically, we prove that if the geometric realization of $G$ has constant doubling dimension, then for a given $s-t$ pair, we only need to probe the weights on $((\log n) / \epsilon)^{O(1)}$ edges in $G'$ in order to obtain a $(1 + \epsilon)$-approximation to the $s-t$ distance in $G'$. We also demonstrate experimentally that this result is pessimistic -- one can even obtain a short path in $G'$ with a small number of probes to $G'$.
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
Optimization by Decoded Quantum Interferometry
Stephen Jordan
Noah Shutty
Mary Wootters
Alexander Schmidhuber
Robbie King
Sergei Isakov
arXiv:2408.08292 (2024)
Preview abstract
We introduce Decoded Quantum Interferometry (DQI), a quantum algorithm for reducing classical optimization problems to classical decoding problems by exploiting structure in the Fourier spectrum of the objective function. DQI reduces sparse max-XORSAT to decoding LDPC codes, which can be decoded using powerful classical algorithms such as belief propagation. As an initial benchmark, we compare DQI using belief propagation decoding against classical optimization via simulated annealing. In this setting we identify a family of max-XORSAT instances where DQI achieves a better approximation ratio on average than simulated annealing, although not better than specialized classical algorithms tailored to those instances. We also analyze a combinatorial optimization problem corresponding to finding polynomials that intersect the maximum number of points. There, DQI efficiently achieves a better approximation ratio than any polynomial-time classical algorithm known to us, thus realizing an apparent exponential quantum speedup. Finally, we show that the problem defined by Yamakawa and Zhandry in order to prove an exponential separation between quantum and classical query complexity is a special case of the optimization problem efficiently solved by DQI.
View details
Preview abstract
Given a training data-set $\mathcal{S}$, and a reference data-set $\mathcal{T}$, we design a simple and efficient algorithm to reweigh the loss function such that the limiting distribution of the neural network weights that result from training on $\mathcal{S}$ approaches the limiting distribution that would have resulted by training on $\mathcal{T}$. Such reweighing can be used to correct for Train-Test distribution shift, when we don't have access to the labels of $\mathcal{T}$. It can also be used to perform (soft) multi-criteria optimization on neural nets, when we have access to the labels of $\mathcal{T}$, but $\mathcal{S}$ and $\mathcal{T}$ have few common points.
As a motivating application, we train a graph neural net to recognize small molecule binders to MNK2 (a MAP Kinase, responsible for cell signaling) which are non-binders to MNK1 (a very similar protein), even in the absence of training data common to both data-sets. We are able to tune the reweighing parameters so that overall change in holdout loss is negligible, but the selectivity, i.e., the fraction of top 100 MNK2 binders that are MNK1 non-binders, increases from 54\% to 95\%, as a result of our reweighing.
We expect the algorithm to be applicable in other settings as well, since we prove that when the metric entropy of the input data-sets is bounded, our random sampling based greedy algorithm outputs a close to optimal reweighing, i.e., the two invariant distributions of network weights will be provably close in total variation distance.
View details
"We Need Structured Output": Towards User-centered Constraints on Large Language Model Output
Michael Xieyang Liu
Frederick Liu
Alex Fiannaca
Terry Koo
In Extended Abstract in ACM CHI Conference on Human Factors in Computing Systems (CHI EA '24), ACM (2024), pp. 9 (to appear)
Preview abstract
Large language models can produce creative and diverse responses. However, to integrate them into current developer workflows, it is essential to constrain their outputs to follow specific formats or standards. In this work, we surveyed 51 experienced industry professionals to understand the range of scenarios and motivations driving the need for output constraints from a user-centered perspective. We identified 134 concrete use cases for constraints at two levels: low-level, which ensures the output adhere to a structured format and an appropriate length, and high-level, which requires the output to follow semantic and stylistic guidelines without hallucination. Critically, applying output constraints could not only streamline the currently repetitive process of developing, testing, and integrating LLM prompts for developers, but also enhance the user experience of LLM-powered features and applications. We conclude with a discussion on user preferences and needs towards articulating intended constraints for LLMs, alongside an initial design for a constraint prototyping tool.
View details
Learning from Label Proportions: Bootstrapping Supervised Learners via Belief Propagation
Shreyas Havaldar
The Twelfth International Conference on Learning Representations (ICLR) (2024)
Preview abstract
Learning from Label Proportions (LLP) is a learning problem where only aggregate level labels are available for groups of instances, called bags, during training, and the aim is to get the best performance at the instance-level on the test data. This setting arises in domains like advertising and medicine due to privacy considerations. We propose a novel algorithmic framework for this problem that iteratively performs two main steps. For the first step (Pseudo Labeling) in every iteration, we define a Gibbs distribution over binary instance labels that incorporates a) covariate information through the constraint that instances with similar covariates should have similar labels and b) the bag level aggregated label. We then use Belief Propagation (BP) to marginalize the Gibbs distribution to obtain pseudo labels. In the second step (Embedding Refinement), we use the pseudo labels to provide supervision for a learner that yields a better embedding. Further, we iterate on the two steps again by using the second step's embeddings as new covariates for the next iteration. In the final iteration, a classifier is trained using the pseudo labels. Our algorithm displays strong gains against several SOTA baselines for the LLP Binary Classification problem on various dataset types - Small Tabular, Large Tabular and Images. We achieve these improvements with minimal computational overhead above standard supervised learning due to Belief Propagation, for large bag sizes, even for a million samples.
View details
Pathfinder: High-Resolution Control-Flow Attacks with Conditional Branch Predictor
Andrew Kwong
Archit Agarwal
Christina Garman
Daniel Genkin
Dean Tullsen
Deian Stefan
Hosein Yavarzadeh
Max Christman
Mohammadkazem Taram
International Conference on Architectural Support for Programming Languages and Operating Systems, ACM (2024)
Preview abstract
This paper presents novel attack primitives that provide adversaries with the ability to read and write the path history register (PHR) and the prediction history tables (PHTs) of the conditional branch predictor in modern Intel CPUs. These primitives enable us to recover the recent control flow (the last 194 taken branches) and, in most cases, a nearly unlimited control flow history of any victim program. Additionally, we present a tool that transforms the PHR into an unambiguous control flow graph, encompassing the complete history of every branch. This work provides case studies demonstrating the practical impact of novel reading and writing/poisoning primitives. It includes examples of poisoning AES to obtain intermediate values and consequently recover the secret AES key, as well as recovering a secret image by capturing the complete control flow of libjpeg routines. Furthermore, we demonstrate that these attack primitives are effective across virtually all protection boundaries and remain functional in the presence of all recent control-flow mitigations from Intel.
View details
