Pushmeet Kohli
Pushmeet Kohli is the Chief Scientist of Google Cloud and VP of Research at Google DeepMind.
Pushmeet founded and leads Google DeepMind's AI for Science and Strategic Initiatives Unit, which develops AI systems to pursue breakthroughs in biology, materials, fusion, math, software, and cybersecurity.
Key achievements include the Nobel Prize-winning AlphaFold model for protein structure prediction, one of the most cited AI biology papers (40,000+ citations) that led to the DeepMind spin-off, Isomorphic Labs. The team has produced a number of state of the art models and agents including AlphaGenome, AlphaEarth, AlphaProof, AlphaProteo, SynthID, AlphaCode. Kohli's team also pioneered large language model-based AI agents like AlphaEvolve and AI Co-Scientist, leading to discoveries in math and biology.
Before DeepMind, he spent a decade at Microsoft Research, becoming Director of Research for the Cognition group. His doctoral thesis received the Sullivan Doctoral Thesis Award. In 2023, TIME recognized him as one of the 100 most influential people in AI.
Authored Publications
Sort By
Automated loss of pulse detection on a commercial smartwatch
Kamal Shah
Yiwen Chen
Anthony Stange
Lawrence Cai
Matt Wimmer
Pramod Rudrapatna
Shelten Yuen
Anupam Pathak
Shwetak Patel
Mark Malhotra
Marc Stogaitis
Jeanie Phan
Ali Connell
Jim Taylor
Jacqueline Shreibati
Daniel McDuff
Tajinder Gadh
Jake Sunshine
Nature, 642 (2025), pp. 174-181
Preview abstract
Out-of-hospital cardiac arrest is a time-sensitive emergency that requires prompt identification and intervention: sudden, unwitnessed cardiac arrest is nearly unsurvivable. A cardinal sign of cardiac arrest is sudden loss of pulse. Automated biosensor detection of unwitnessed cardiac arrest, and dispatch of medical assistance, may improve survivability given the substantial prognostic role of time, but only if the false-positive burden on public emergency medical systems is minimized. Here we show that a multimodal, machine learning-based algorithm on a smartwatch can reach performance thresholds making it deployable at a societal scale. First, using photoplethysmography, we show that wearable photoplethysmography measurements of peripheral pulselessness (induced through an arterial occlusion model) manifest similarly to pulselessness caused by a common cardiac arrest arrhythmia, ventricular fibrillation. On the basis of the similarity of the photoplethysmography signal (from ventricular fibrillation or arterial occlusion), we developed and validated a loss of pulse detection algorithm using data from peripheral pulselessness and free-living conditions. Following its development, we evaluated the end-to-end algorithm prospectively: there was 1 unintentional emergency call per 21.67 user-years across two prospective studies; the sensitivity was 67.23% (95% confidence interval of 64.32% to 70.05%) in a prospective arterial occlusion cardiac arrest simulation model. These results indicate an opportunity, deployable at scale, for wearable-based detection of sudden loss of pulse while minimizing societal costs of excess false detections.
View details
Olympiad-Level Formal Mathematical Reasoning with Reinforcement Learning
Thomas Hubert
Rishi Mehta
Laurent Sartran
Miklós Z. Horváth
Eric Wieser
Aja Huang
Julian Schrittwieser
Yannick Schroecker
Hussain Masoom
Ottavia Bertolli
Tom Zahavy
Amol Mandhane
Jessica Yung
Iuliya Beloshapka
Borja Ibarz
Vivek Veeriah
Lei Yu
Oliver Nash
Paul Lezeau
Salvatore Mercuri
Calle Sönne
Bhavik Mehta
Alex Davies
Daniel Zheng
Yin Li
Ingrid von Glehn
Mark Rowland
Samuel Albanie
Ameya Velingker
Simon Schmitt
Henryk Michalewski
Nicolas Sonnerat
Demis Hassabis
David Silver
Nature (2025)
Preview abstract
A long-standing goal of artificial intelligence is to build systems capable of complex reasoning in vast domains, a task epitomized by mathematics with its boundless concepts and demand for rigorous proof. Recent AI systems, often reliant on human data, typically lack the formal verification necessary to guarantee correctness. By contrast, formal languages such as Lean1 offer an interactive environment that grounds reasoning, and reinforcement learning (RL) provides a mechanism for learning in such environments. We present AlphaProof, an AlphaZero-inspired2 agent that learns to find formal proofs through RL by training on millions of auto-formalized problems. For the most difficult problems, it uses Test-Time RL, a method of generating and learning from millions of related problem variants at inference time to enable deep, problem-specific adaptation. AlphaProof substantially improves state-of-the-art results on historical mathematics competition problems. At the 2024 IMO competition, our AI system, with AlphaProof as its core reasoning engine, solved three out of the five non-geometry problems, including the competition’s most difficult problem. Combined with AlphaGeometry 23, this performance, achieved with multi-day computation, resulted in reaching a score equivalent to that of a silver medallist, marking the first time an AI system achieved any medal-level performance. Our work demonstrates that learning at scale from grounded experience produces agents with complex mathematical reasoning strategies, paving the way for a reliable AI tool in complex mathematical problem-solving.
View details
Scaling Wearable Foundation Models
Girish Narayanswamy
Kumar Ayush
Yuzhe Yang
Orson Xu
Shun Liao
Shyam Tailor
Jake Sunshine
Tim Althoff
Shrikanth (Shri) Narayanan
Jiening Zhan
Mark Malhotra
Shwetak Patel
Samy Abdel-Ghaffar
Daniel McDuff
2025
Preview abstract
Wearable sensors have become ubiquitous thanks to a variety of health tracking features. The resulting continuous and longitudinal measurements from everyday life generate large volumes of data. However, making sense of these observations for scientific and actionable insights is non-trivial. Inspired by the empirical success of generative modeling, where large neural networks learn powerful representations from vast amounts of text, image, video, or audio data, we investigate the scaling properties of wearable sensor foundation models across compute, data, and model size. Using a dataset of up to 40 million hours of in-situ heart rate, heart rate variability, accelerometer, electrodermal activity, skin temperature, and altimeter per-minute data from over 165,000 people, we create LSM, a multimodal foundation model built on the largest wearable-signals dataset with the most extensive range of sensor modalities to date. Our results establish the scaling laws of LSM for tasks such as imputation, interpolation and extrapolation across both time and sensor modalities. Moreover, we highlight how LSM enables sample-efficient downstream learning for tasks including exercise and activity recognition.
View details
Generative models improve fairness of medical classifiers under distribution shifts
Ira Ktena
Olivia Wiles
Isabela Albuquerque
Sylvestre-Alvise Rebuffi
Ryutaro Tanno
Danielle Belgrave
Taylan Cemgil
Nature Medicine (2024)
Preview abstract
Domain generalization is a ubiquitous challenge for machine learning in healthcare. Model performance in real-world conditions might be lower than expected because of discrepancies between the data encountered during deployment and development. Underrepresentation of some groups or conditions during model development is a common cause of this phenomenon. This challenge is often not readily addressed by targeted data acquisition and ‘labeling’ by expert clinicians, which can be prohibitively expensive or practically impossible because of the rarity of conditions or the available clinical expertise. We hypothesize that advances in generative artificial intelligence can help mitigate this unmet need in a steerable fashion, enriching our training dataset with synthetic examples that address shortfalls of underrepresented conditions or subgroups. We show that diffusion models can automatically learn realistic augmentations from data in a label-efficient manner. We demonstrate that learned augmentations make models more robust and statistically fair in-distribution and out of distribution. To evaluate the generality of our approach, we studied three distinct medical imaging contexts of varying difficulty: (1) histopathology, (2) chest X-ray and (3) dermatology images. Complementing real samples with synthetic ones improved the robustness of models in all three medical tasks and increased fairness by improving the accuracy of clinical diagnosis within underrepresented groups, especially out of distribution.
View details
Consensus, dissensus and synergy between clinicians and specialist foundation models in radiology report generation
Ryutaro Tanno
David Barrett
Sumedh Ghaisas
Sumanth Dathathri
Abi See
Johannes Welbl
Karan Singhal
Rhys May
Roy Lee
SiWai Man
Zahra Ahmed
Sara Mahdavi
Joelle Barral
Ali Eslami
Danielle Belgrave
Shravya Shetty
Po-Sen Huang
Ira Ktena
Arxiv (2023)
Preview abstract
Radiology reports are an instrumental part of modern medicine, informing key clinical decisions such as diagnosis and treatment. The worldwide shortage of radiologists, however, restricts access to expert care and imposes heavy workloads, contributing to avoidable errors and delays in report delivery. While recent progress in automated report generation with vision-language models offer clear potential in ameliorating the situation, the path to real-world adoption has been stymied by the challenge of evaluating the clinical quality of AI-generated reports. In this study, we build a state-of-the-art report generation system for chest radiographs, Flamingo-CXR, by fine-tuning a well-known vision-language foundation model on radiology data. To evaluate the quality of the AI-generated reports, a group of 16 certified radiologists provide detailed evaluations of AI-generated and human written reports for chest X-rays from an intensive care setting in the United States and an inpatient setting in India. At least one radiologist (out of two per case) preferred the AI report to the ground truth report in over 60% of cases for both datasets. Amongst the subset of AI-generated reports that contain errors, the most frequently cited reasons were related to the location and finding, whereas for human written reports, most mistakes were related to severity and finding. This disparity suggested potential complementarity between our AI system and human experts, prompting us to develop an assistive scenario in which Flamingo-CXR generates a first-draft report, which is subsequently revised by a clinician. This is the first demonstration of clinician-AI collaboration for report writing, and the resultant reports are assessed to be equivalent or preferred by at least one radiologist to reports written by experts alone in 80% of in-patient cases and 60% of intensive care cases.
View details
Enhancing diagnostic accuracy of medical AI systems via selective deferral to clinicians
Dj Dvijotham
Jim Winkens
Melih Barsbey
Sumedh Ghaisas
Robert Stanforth
Nick Pawlowski
Patricia Strachan
Zahra Ahmed
Yoram Bachrach
Laura Culp
Mayank Daswani
Jan Freyberg
Christopher Kelly
Atilla Kiraly
Timo Kohlberger
Scott Mayer McKinney
Basil Mustafa
Krzysztof Geras
Jan Witowski
Zhi Zhen Qin
Jacob Creswell
Shravya Shetty
Terry Spitz
Taylan Cemgil
Nature Medicine (2023)
Preview abstract
AI systems trained using deep learning have been shown to achieve expert-level identification of diseases in multiple medical imaging settings1,2. While these results are impressive, they don’t accurately reflect the impact of deployment of such systems in a clinical context. Due to the safety-critical nature of this domain and the fact that AI systems are not perfect and can make inaccurate assessments, they are predominantly deployed as assistive tools for clinical experts3. Although clinicians routinely discuss the diagnostic nuances of medical images with each other, weighing human diagnostic confidence against that of an AI system remains a major unsolved barrier to collaborative decision-making4. Furthermore, it has been observed that diagnostic AI models have complementary strengths and weaknesses compared to clinical experts. Yet, complementarity and the assessment of relative confidence between the members of a diagnostic team has remained largely unexploited in how AI systems are currently used in medical settings5.
In this paper, we study the behavior of a team composed of diagnostic AI model(s) and clinician(s) in diagnosing disease. To go beyond the performance level of a standalone AI system, we develop a novel selective deferral algorithm that can learn to decide when to rely on a diagnostic AI model and when to defer to a clinical expert. Using this algorithm, we demonstrate that the composite AI+human system has enhanced accuracy (both sensitivity and specificity) relative to a human-only or an AI-only baseline. We decouple the development of the deferral AI model from training of the underlying diagnostic AI model(s). Development of the deferral AI model only requires i) the predictions of a model(s) on a tuning set of medical images (separate from the diagnostic AI models’ training data), ii) the diagnoses made by clinicians on these images and iii) the ground truth disease labels corresponding to those images.
Our extensive analysis shows that the selective deferral (SD) system exceeds the performance of either clinicians or AI alone in multiple clinical settings: breast and lung cancer screening. For breast cancer screening, double-reading with arbitration (two readers interpreting each mammogram invoking an arbitrator if needed) is a “gold standard” for performance, never previously exceeded using AI6. The SD system exceeds the accuracy of double-reading with arbitration in a large representative UK screening program (25% reduction in false positives despite equivalent true-positive detection and 66% reduction in the requirement for clinicians to read an image), as well as exceeding the performance of a standalone state-of-art AI system (40% reduction in false positives with equivalent detection of true positives). In a large US dataset the SD system exceeds the accuracy of single-reading by board-certified radiologists and a standalone state-of-art AI system (32% reduction in false positives despite equivalent detection of true positives and 55% reduction in the clinician workload required). The SD system further outperforms both clinical experts alone, and AI alone for the detection of lung cancer in low-dose Computed Tomography images from a large national screening study, with 11% reduction in false positives while maintaining sensitivity given 93% reduction in clinician workload required. Furthermore, the SD system allows controllable trade-offs between sensitivity and specificity and can be tuned to target either specificity or sensitivity as desired for a particular clinical application, or a combination of both.
The system generalizes to multiple distribution shifts, retaining superiority to both the AI system alone and human experts alone. We demonstrate that the SD system retains performance gains even on clinicians not present in the training data for the deferral AI. Furthermore, we test the SD system on a new population where the standalone AI system’s performance significantly degrades. We showcase the few-shot adaptation capability of the SD system by demonstrating that the SD system can obtain superiority to both the standalone AI system and the clinician on the new population after being trained on only 40 cases from the new population.
Our comprehensive assessment demonstrates that a selective deferral system could significantly improve clinical outcomes in multiple medical imaging applications, paving the way for higher performance clinical AI systems that can leverage the complementarity between clinical experts and medical AI tools.
View details
Reinforced Genetic Algorithm Learning for Optimizing Computation Graphs
Aditya Paliwal
Felix Gimeno
Vinod Gopal Nair
Yujia Li
Miles Lubin
International Conference on Learning Representations (ICLR) (2020)
Preview abstract
We present a deep reinforcement learning approach to minimizing the execution cost of neural network computation graphs in an optimizing compiler. Unlike earlier learning-based works that require training the optimizer on the same graph to be optimized, we propose a learning approach that trains an optimizer offline and then generalizes to previously unseen graphs without further training. This allows our approach to produce high-quality execution decisions on real-world TensorFlow graphs in seconds instead of hours. We consider two optimization tasks for computation graphs: minimizing running time and peak memory usage. In comparison to an extensive set of baselines, our approach achieves significant improvements over classical and other learning-based methods on these two tasks.
View details
Learning Transferable Graph Exploration
Hanjun Dai
Yujia Li
Chenglong Wang
Rishabh Singh
Po-Sen Huang
Neural Processing Information Systems (NeurIPS) (2019)
Preview abstract
This paper considers the problem of efficient exploration of unseen environments, a key challenge in AI. We propose a `learning to explore' framework where we learn a policy from a distribution of environments. At test time, presented with an unseen environment from the same distribution, the policy aims to generalize the exploration strategy to visit the maximum number of unique states in a limited number of steps. We particularly focus on environments with graph-structured state-spaces that are encountered in many important real-world applications like software testing and map building. We formulate this task as a reinforcement learning problem where the `exploration' agent is rewarded for transitioning to previously unseen environment states and employ a graph-structured memory to encode the agent's past trajectory. Experimental results demonstrate that our approach is extremely effective for exploration of spatial maps; and when applied on the challenging problems of coverage-guided software-testing of domain-specific programs and real-world mobile applications, it outperforms methods that have been hand-engineered by human experts.
View details
Relational inductive biases, deep learning, and graph networks
Peter Battaglia
Jessica Blake Chandler Hamrick
Victor Bapst
Alvaro Sanchez
Vinicius Zambaldi
Mateusz Malinowski
Andrea Tacchetti
David Raposo
Adam Santoro
Ryan Faulkner
Caglar Gulcehre
Francis Song
Andy Ballard
Justin Gilmer
Ashish Vaswani
Kelsey Allen
Charles Nash
Victoria Jayne Langston
Chris Dyer
Nicolas Heess
Daan Wierstra
Matt Botvinick
Yujia Li
Razvan Pascanu
arXiv (2018)
Preview abstract
The purpose of this paper is to explore relational inductive biases in modern AI, especially
deep learning, describing a rough taxonomy of existing approaches, and introducing a common
mathematical framework for expressing and unifying various approaches. The key theme running through this work is structure—how the world is structured, and how the structure of different computational strategies determines their strengths and weaknesses.
View details
Preview abstract
We present a reinforcement learning framework, called Programmatically Interpretable Reinforcement Learning (PIRL), that is designed to generate interpretable and verifiable agent policies. Unlike the popular Deep Reinforcement Learning (DRL) paradigm, which represents policies by neural networks, PIRL represents policies using a high-level, domain-specific programming language. Such programmatic policies have the benefits of being more easily interpreted than neural networks, and being amenable to verification by symbolic methods. We propose a new method, called Neurally Directed Program Search (NDPS), for solving the challenging nonsmooth optimization problem of finding a programmatic policy with maximal reward. NDPS works by first learning a neural policy network using DRL, and then performing a local search over programmatic policies that seeks to minimize a distance from this neural “oracle”. We evaluate NDPS on the task of learning to drive a simulated car in the TORCS car-racing environment. We demonstrate that NDPS is able to discover human-readable policies that pass some significant performance bars. We also show that PIRL policies can have smoother trajectories, and can be more easily transferred to environments not encountered during training, than corresponding policies discovered by DRL.
View details
