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 10822 publications
mmMUSE: An mmWave-based Motion-resilient Universal Speech Enhancement System
Chenming He
Yanyong Zhang
Kai Wang
Dequan Wang
Lingyu Wang
the Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies (IMWUT), ACM (2026) (to appear)
Preview abstract
Voice-based smart systems can greatly enhance user experiences by allowing higher-quality interactions through better voice perception. Speech enhancement can benefit such systems by isolating noise from speech. Recently, integrating millimeter-wave (mmWave) with audio for speech perception has gained increasing attention due to microphones' limitations in noisy environments. However, mmWave-based vocal extraction is severely affected by motion, which disperses vocal signals across ranges and introduces distortions. In this paper, we propose an mmWave-based motion-resilient universal speech enhancement system called mmMUSE, which fuses mmWave and audio signals. To mitigate motion interference, we develop a Doppler-based method for motion-robust vocal signal extraction. Moreover, by introducing the Vocal-Noise-Ratio metric to assess the prominence of vocal signals from mmWave, we achieve real-time voice activity detection that gains 3.81 dB of SISDR in noisy speeches. Additionally, we design a two-stage complex-valued network that includes an attention-based fusion network for cross-modal complementing and a time-frequency masking network for correcting amplitude and phase of speech to isolate noises.
Using mmWave and audio datasets from 46 participants, mmMUSE outperforms the state-of-the-art speech enhancement models, achieving an average SISDR improvement of 3.12 dB. Additionally, mmMUSE achieves SISDR improvements of 16.51 dB, 17.93 dB, 14.93 dB, and 18.95 dB in controlled environments involving intense noise, extensive motion, multiple speakers, and various obstructive materials, respectively. Finally, we evaluate mmMUSE in real-world scenarios including running, public spaces, and driving, maintaining a word error rate (WER) below 10%.
View details
Preview abstract
For many practical applications of quantum computing, the slowest and most costly steps involve coherently accessing classical data. We help address this challenge by applying mass production techniques, which can sometimes allow us to perform operations many times in parallel for a cost that is comparable to a single execution[1-3]. We combine existing mass-production results with modern approaches for loading classical data using ``quantum read-only memory.'' We show that quantum mass production techniques offer no benefit when we consider a cost model that focuses purely on the number of non-Clifford gates. However, analyzing the constant factors in a more nuanced cost model, we find that it may be possible to obtain a reduction in cost of an order or magnitude or more for a variety reasonably-sized fault-tolerant quantum algorithms. We present several applications of quantum mass-production techniques beyond naive parallelization, including a strategy for reducing the cost of serial calls to the same data loading step.
View details
Preview abstract
AI coding assistants are rapidly becoming integral to modern software development. A key challenge in this space is the continual need to migrate and modernize codebases in response to evolving software ecosystems. Traditionally, such migrations have relied on rule-based systems and human intervention. With the advent of powerful large language models (LLMs), AI-driven agentic frameworks offer a promising alternative—but their effectiveness remains underexplored. In this paper, we introduce FreshBrew, a novel benchmark for evaluating AI-based agentic frameworks on project-level Java migrations. We benchmark several such frameworks, powered by state-of-the-art LLMs, and compare their performance against established rule-based tools. Our evaluation of AI agents on this benchmark of 228 repositories shows that the top-performing model, Gemini 2.5 Flash, can successfully migrate 56.5% of projects to JDK 17. Our empirical analysis reveals novel insights into the critical strengths and limitations of current agentic approaches, offering actionable insights into their real-world applicability. By releasing FreshBrew publicly upon acceptance, we aim to facilitate rigorous, reproducible evaluation and catalyze progress in AI-driven codebase modernization.
View details
Dynamical-generative downscaling of climate model ensembles
Tapio Schneider
John Anderson
Fei Sha
Proceedings of the National Academy of Sciences, 122 (2025), e2420288122
Preview abstract
Regional high-resolution climate projections are crucial for many applications, such as agriculture, hydrology, and natural hazard risk assessment. Dynamical downscaling, the state-of-the-art method to produce localized future climate information, involves running a regional climate model (RCM) driven by an Earth System Model (ESM), but it is too computationally expensive to apply to large climate projection ensembles. We propose an approach combining dynamical downscaling with generative AI to reduce the cost and improve the uncertainty estimates of downscaled climate projections. In our framework, an RCM dynamically downscales ESM output to an intermediate resolution, followed by a generative diffusion model that further refines the resolution to the target scale. This approach leverages the generalizability of physics-based models and the sampling efficiency of diffusion models, enabling the downscaling of large multimodel ensembles. We evaluate our method against dynamically downscaled climate projections from the Coupled Model Intercomparison Project 6 (CMIP6) ensemble. Our results demonstrate its ability to provide more accurate uncertainty bounds on future regional climate than alternatives such as dynamical downscaling of smaller ensembles, or traditional empirical statistical downscaling methods. We also show that dynamical-generative downscaling results in significantly lower errors than popular statistical downscaling techniques, and captures more accurately the spectra, tail dependence, and multivariate correlations of meteorological fields. These characteristics make the dynamical-generative framework a flexible, accurate, and efficient way to downscale large ensembles of climate projections, currently out of reach for pure dynamical downscaling.
View details
On Fisher Consistency of Surrogate Losses for Optimal Dynamic Treatment Regimes with Multiple Categorical Treatments per Stage
Nilanjana Laha
Nilson Chapagain
Victoria Cicherski
Aarón Sonabend
arxiv (2025) (to appear)
Preview abstract
Patients with chronic diseases often receive treatments at multiple time points, or stages. Our goal is to learn the optimal dynamic treatment regime (DTR) from longitudinal patient data. When both the number of stages and the number of treatment levels per stage are arbitrary, estimating the optimal DTR reduces to a sequential, weighted, multiclass classification problem (Kosorok and Laber, 2019). In this paper, we aim to solve this classification problem simultaneously across all stages using Fisher consistent surrogate losses. Although computationally feasible Fisher consistent surrogates exist in special cases, e.g., the binary treatment setting, a unified theory of Fisher consistency remains largely unexplored. We establish necessary and sufficient conditions for DTR Fisher consistency within the class of non-negative, stagewise separable surrogate losses. To our knowledge, this is the first result in the DTR literature to provide necessary conditions for Fisher consistency within a non-trivial surrogate class. Furthermore, we show that many convex surrogate losses fail to be Fisher consistent for the DTR classification problem, and we formally establish this inconsistency for smooth, permutation equivariant, and relative-margin-based convex losses. Building on this, we propose SDSS (Simultaneous Direct Search with Surrogates), which uses smooth, non-concave surrogate losses to learn the optimal DTR. We develop a computationally efficient, gradient-based algorithm for SDSS. When the optimization error is small, we establish a sharp upper bound on SDSS's regret decay rate. We evaluate the numerical performance of SDSS through simulations and demonstrate its real-world applicability by estimating optimal fluid resuscitation strategies for severe septic patients using electronic health record data.
View details
Mastering Multiple-Expert Routing: Realizable H-Consistency and Strong Guarantees for Learning to Defer
Anqi Mao
Proceedings of the 42nd International Conference on Machine Learning (ICML 2025)
Preview abstract
The problem of learning to defer with multiple experts consists of optimally assigning input instances to experts, balancing the trade-off between their accuracy and computational cost. This is a critical challenge in natural language generation, but also in other fields such as image processing, and medical diagnostics. Recent studies have proposed surrogate loss functions to optimize deferral, but challenges remain in ensuring their consistency properties. This paper introduces novel surrogate loss functions and efficient algorithms with strong theoretical learning guarantees. We address open questions regarding realizable $H$-consistency, $H$-consistency bounds, and Bayes-consistency for both single-stage (jointly learning predictor and deferral function) and two-stage (learning only the deferral function with a fixed expert) learning scenarios. For single-stage deferral, we introduce a family of new realizable $H$-consistent surrogate losses and further prove $H$-consistency for a selected member. For two-stage deferral, we derive new surrogate losses that achieve realizable $H$-consistency, $H$-consistency bounds, and Bayes-consistency for the two-expert scenario and, under natural assumptions, multiple-expert scenario. Additionally, we provide enhanced theoretical guarantees under low-noise assumptions for both scenarios. Finally, we report the results of experiments using our proposed surrogate losses, comparing their performance against existing baselines.
View details
AfriMed-QA: A Pan-African Multi-Specialty Medical Question-Answering Benchmark Dataset
Tobi Olatunji
Abraham Toluwase Owodunni
Charles Nimo
Jennifer Orisakwe
Henok Biadglign Ademtew
Chris Fourie
Foutse Yuehgoh
Stephen Moore
Mardhiyah Sanni
Emmanuel Ayodele
Timothy Faniran
Bonaventure F. P. Dossou
Fola Omofoye
Wendy Kinara
Tassallah Abdullahi
Michael Best
2025
Preview abstract
Recent advancements in large language model (LLM) performance on medical multiple-choice question (MCQ) benchmarks have stimulated significant interest from patients and healthcare providers globally. Particularly in low- and middle-income countries (LMICs) facing acute physician shortages and lack of specialists, LLMs offer a potentially scalable pathway to enhance healthcare access and reduce costs. However, LLM training data is sourced from predominantly Western text, existing benchmarks are predominantly Western-centric, limited to MCQs, and focused on a narrow range of clinical specialties, raising concerns about their applicability in the Global South, particularly across Africa where localized medical knowledge and linguistic diversity are often underrepresented. In this work, we introduce AfriMed-QA, the first large-scale multi-specialty Pan-African medical Question-Answer (QA) dataset designed to evaluate and develop equitable and effective LLMs for African healthcare. It contains 3,000 multiple-choice professional medical exam questions with answers and rationale, 1,500 short answer questions (SAQ) with long-from answers, and 5,500 consumer queries, sourced from over 60 medical schools across 15 countries, covering 32 medical specialties. We further rigorously evaluate multiple open, closed, general, and biomedical LLMs across multiple axes including accuracy, consistency, factuality, bias, potential for harm, local geographic relevance, medical reasoning, and recall. We believe this dataset provides a valuable resource for practical application of large language models in African healthcare and enhances the geographical diversity of health-LLM benchmark datasets.
View details
Probing non-equilibrium topological order on a quantum processor
Melissa Will
Tyler Cochran
Bernhard Jobst
Norhan Eassa
Michael Knap
Adam Gammon-Smith
Frank Pollmann
Nature, 645 (2025), 348–353
Preview abstract
Out-of-equilibrium phases in many-body systems constitute a new paradigm in quantum matter—they exhibit dynamical properties that may otherwise be forbidden by equilibrium thermodynamics. Among these non-equilibrium phases are periodically driven (Floquet) systems, which are generically difficult to simulate classically because of their high entanglement. Here we realize a Floquet topologically ordered state on an array of superconducting qubits. We image the characteristic dynamics of its chiral edge modes and characterize its emergent anyonic excitations. Devising an interferometric algorithm allows us to introduce and measure a bulk topological invariant to probe the dynamical transmutation of anyons for system sizes up to 58 qubits. Our work demonstrates that quantum processors can provide key insights into the thus-far largely unexplored landscape of highly entangled non-equilibrium phases of matter.
View details
Preview abstract
Several resource allocation settings involve agents with unequal entitlements represented by weights. We analyze weighted fair division from an asymptotic perspective: if m items are divided among n agents whose utilities are independently sampled from a probability distribution, when is it likely that a fair allocation exist? We show that if the ratio between the weights is bounded, a weighted envy-free allocation exists with high probability provided that m = Ω(n log n/ log log n), generalizing a prior unweighted result. For weighted proportionality, we establish a sharp threshold of m = n/(1 − μ) for the transition from non-existence to existence, where μ ∈ (0, 1) denotes the mean of the distribution. In addition, we prove that for two agents, a weighted envy-free (and weighted proportional) allocation is likely to exist if m = ω(√r), where r denotes the ratio between the two weights.
View details
Understanding challenges to the validity of disaggregated evaluations for algorithmic fairness
Chirag Nagpal
David Madras
Vishwali Mhasawade
Olawale Salaudeen
Shannon Sequeira
Santiago Arciniegas
Lillian Sung
Nnamdi Ezeanochie
Heather Cole-Lewis
Sanmi Koyejo
Proceedings of the 2025 Conference on Neural Information Processing Systems (NeurIPS) (2025)
Preview abstract
Disaggregated evaluation across subgroups is critical for assessing the fairness of machine learning models, but its uncritical use can mislead practitioners. We show that equal performance across subgroups is an unreliable measure of fairness when data are representative of the relevant populations but reflective of real-world disparities. Furthermore, when data are not representative due to selection bias, both disaggregated evaluation and alternative approaches based on conditional independence testing may be invalid without explicit assumptions regarding the bias mechanism. We use causal graphical models to characterize fairness properties and metric stability across subgroups under different data generating processes. Our framework suggests complementing disaggregated evaluations with explicit causal assumptions and analysis to control for confounding and distribution shift, including conditional independence testing and weighted performance estimation. These findings have broad implications for how practitioners design and interpret model assessments given the ubiquity of disaggregated evaluation.
View details
Preview abstract
Although sound information extraction appear distinct across spectrum of sound classes and technologies, all inherently involve creating some form of "embedding"—be it discrete as in textual tokens or continuous vectors—to encapsulate relevant information from the audio signal for downstream utilization. This unifying framework allows us to re-evaluate sound information extraction by researching the optimality of current task-specific representations, the quality headroom and the potential for a single, robust sound embedding to generalize across diverse applications and sound types. To expedite research in these directions, a standardized evaluation benchmark is indispensable, mirroring the established benchmarks in text and image domains. We present the Massive Sound Embedding Benchmark (MSEB) to serve this purpose. MSEB encompasses realistic tasks and datasets that reflect practical applications across diverse technologies and sound categories. Initial experimental findings indicate substantial headroom for enhancing prevalent information extraction methodologies. We encourage the sound processing community to contribute data and tasks to MSEB and employ it to assess their algorithms for improved overall sound encoding.
View details
Preview abstract
We study MaxCut on 3-regular graphs of minimum girth $g$ for various $g$'s. We obtain new lower bounds on the maximum cut achievable in such graphs by analyzing the Quantum Approximate Optimization Algorithm (QAOA). For $g\geq14$, at depth $p \geq6$, the QAOA improves on previously known lower bounds. The bounds are established through classical numerical analysis of the QAOA's expected performance. This analysis does not produce the actual cuts but establishes their existence. When implemented on a quantum computer, the QAOA provides an efficient algorithm for finding such cuts, using a constant-depth quantum circuit. To our knowledge, this gives an exponential speedup over the best known classical algorithm guaranteed to achieve cuts of this size on graphs of this girth. We also apply the QAOA to the Maximum Independent Set problem on the same class of graphs.
View details
Preview abstract
Differential privacy can be achieved in a distributed manner, where multiple parties add independent noise such that their sum protects the overall dataset with differential privacy. A common technique here is for each party to sample their noise from the decomposition of an infinitely divisible distribution. We introduce two novel mechanisms in this setting: 1) the generalized discrete Laplace (GDL) mechanism, whose distribution (which is closed under summation) follows from differences of i.i.d. negative binomial shares, and 2) The multi-scale discrete Laplace (MSDLap) mechanism, which follows the sum of multiple i.i.d. discrete Laplace shares at different scales. The mechanisms can be parameterized to have 𝑂(Δ^3𝑒^{−𝜀}) and 𝑂 (min(Δ^3𝑒^{−𝜀}, Δ^2𝑒^{−2𝜀/3})) MSE, respectively, where the latter bound matches known optimality results. Furthermore, the MSDLap mechanism has the optimal MSE including constants as 𝜀 → ∞. We also show a transformation from the discrete setting to the continuous setting, which allows us to transform both mechanisms to the continuous setting and thereby achieve the optimal 𝑂 (Δ^2𝑒^{−2𝜀/3}) MSE. To our knowledge, these are the first infinitely divisible additive noise mechanisms that achieve order-optimal MSE under pure differential privacy for either the discrete or continuous setting, so our work shows formally there is no separation in utility when query-independent noise adding mechanisms are restricted to infinitely divisible noise. For the continuous setting, our result improves upon Pagh and Stausholm’s Arete distribution which gives an MSE of 𝑂(Δ^2𝑒^{−𝜀/4}) [35]. We apply our results to improve a state of the art multi-message shuffle DP protocol from [3] in the high 𝜀 regime.
View details
Preview abstract
Despite coming up on two decades of network verification research, verification tooling continues to see limited real-world adoption and outages continue to occur. Relying on interviews with network engineers and our own experience as a large network operator, we ask why. These conversations reveal that the culprit is traditional verification's reliance on hand-crafted network models, which leads to issues with coverage, correctness, maintainability, and fidelity, ultimately hindering practical applicability and adoption.
To address this, we call for the research community to embrace "model-free verification" through network emulation. Recent technology advancements – maturation of orchestration infrastructure and vendor-provided container images – make it possible to leverage emulation to obtain a high-fidelity converged dataplane from actual router control plane code, and then apply established dataplane verification techniques to this extracted state. We prototype such a system with open-source components, and present early results showing this approach can accurately verify configurations previously untestable, paving the way for more robust, practical network verification.
View details
Closing the AI generalisation gap by adjusting for dermatology condition distribution differences across clinical settings
Rajeev Rikhye
Aaron Loh
Grace Hong
Margaret Ann Smith
Vijaytha Muralidharan
Doris Wong
Michelle Phung
Nicolas Betancourt
Bradley Fong
Rachna Sahasrabudhe
Khoban Nasim
Alec Eschholz
Basil Mustafa
Jan Freyberg
Terry Spitz
Kat Chou
Peggy Bui
Justin Ko
Steven Lin
The Lancet eBioMedicine (2025)
Preview abstract
Background: Generalisation of artificial intelligence (AI) models to a new setting is challenging. In this study, we seek to understand the robustness of a dermatology (AI) model and whether it generalises from telemedicine cases to a new setting including both patient-submitted photographs (“PAT”) and clinician-taken photographs in-clinic (“CLIN”).
Methods: A retrospective cohort study involving 2500 cases previously unseen by the AI model, including both PAT and CLIN cases, from 22 clinics in the San Francisco Bay Area, spanning November 2015 to January 2021. The primary outcome measure for the AI model and dermatologists was the top-3 accuracy, defined as whether their top 3 differential diagnoses contained the top reference diagnosis from a panel of dermatologists per case.
Findings: The AI performed similarly between PAT and CLIN images (74% top-3 accuracy in CLIN vs. 71% in PAT), however, dermatologists were more accurate in PAT images (79% in CLIN vs. 87% in PAT). We demonstrate that demographic factors were not associated with AI or dermatologist errors; instead several categories of conditions were associated with AI model errors (p < 0.05). Resampling CLIN and PAT to match skin condition distributions to the AI development dataset reduced the observed differences (AI: 84% CLIN vs. 79% PAT; dermatologists: 77% CLIN vs. 89% PAT). We demonstrate a series of steps to close the generalisation gap, requiring progressively more information about the new dataset, ranging from the condition distribution to additional training data for rarer conditions. When using additional training data and testing on the dataset without resampling to match AI development, we observed comparable performance from end-to-end AI model fine tuning (85% in CLIN vs. 83% in PAT) vs. fine tuning solely the classification layer on top of a frozen embedding model (86% in CLIN vs. 84% in PAT).
Interpretation: AI algorithms can be efficiently adapted to new settings without additional training data by recalibrating the existing model, or with targeted data acquisition for rarer conditions and retraining just the final layer.
View details
