I'm a Distinguished Research Scientist at Google DeepMind working on artificial intelligence.
Currently, I work on Gemini's recursive self-improvement capabilities. I also currently co-lead the Gemini Prototyping team. I created the initial prototype (codename 'Neural Google') that led to the creation of Google Search AI Mode, as well as prototypes that led to the adoption of Generative UI across Google.
Prior to working on Gemini, I led the DL:X team at DeepMind (mostly generative models, self-supervised learning, multi-modal large language models), as well as the Quantum Chemistry and Materials team in Science (mostly DFT and a few downstream applications).
See our ICML tutorial for a technical overview of some of that pre-Gemini work. See articles and videos on Quanta, Ars Technica, Stephen Colbert, TEDx and TANK for a non-technical overview of that work. See here and here for a flavour of the team's on-going quantum physics and chemistry research.
Prior to DeepMind, I was a post-doctoral researcher at Microsoft Research Cambridge working with John Winn. I did my PhD at the University of Edinburgh, where I was a Carnegie scholar working with Christopher Williams. During my PhD I was also a visiting researcher at Oxford University working with Andrew Zisserman.
Highlights
-
LLMs -
Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long context, and Next Generation Agentic Capabilities July 2025 -
Multimodal Few-Shot Learning with Frozen Language Models Neural Information Processing Systems (NeurIPS), 2021. -
Computer vision -
Representation Learning Without Labels International Conference on Machine Learning (ICML) tutorial, 2020. -
Self-supervised Video Pretraining Yields Human-Aligned Visual Representations Neural Information Processing Systems (NeurIPS), New Orleans, USA, 2023. -
Data-Efficient Image Recognition with Contrastive Predictive Coding International Conference on Machine Learning (ICML), Vienna, Austria, 2020. -
3D Understanding -
Neural Scene Representation and Rendering Science, 2018.
Extensions of the Generative Query Network: -
Few-shot learning -
Neural Processes and Family International Conference on Machine Learning (ICML), Stockholm, Sweden, 2018.
Overview of the Neural Process family:- Survey (Jha et al.)
-
Creativity -
Unsupervised Doodling and Painting with Improved SPIRAL
-
Robotics -
From Motor Control to Team Play in Simulated Humanoid Football Science Robotics, 2022.
Deployed on real robots (2022):
Precursor work (2017):
All Publications
-
Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long context, and Next Generation Agentic Capabilities -
Advancing Conversational Diagnostic AI with Multimodal Reasoning doi.org/10.1038/s41591-026-04371-0 (Nature Medicine), NM. -
Collaboration Between Clinicians and Vision–Language Models in Radiology Report Generation doi.org/10.1038/s41591-024-03302-1 (Nature Medicine), NM. -
Gemini 1.5: Unlocking Multimodal Understanding Across Millions of Tokens of Context -
Advancing Multimodal Medical Capabilities of Gemini -
Capabilities of Gemini Models in Medicine -
Consensus, Dissensus and Synergy Between Clinicians and Specialist Foundation models in Radiology Report Generation -
Self-supervised Video Pretraining Yields Human-Aligned Visual Representations Neural Information Processing Systems (NeurIPS), New Orleans, USA. -
From Data to Functa: Your Data Point is a Function and You Can Treat It Like One International Conference on Machine Learning (ICML), Baltimore, USA. -
Towards Real Robot Learning in the Wild: A Case Study in Bipedal Locomotion Conference on Robot Learning (CoRL), Auckland, New Zealand. -
Inferring a Continuous Distribution of Atom Coordinates from Cryo-EM Images Using VAEs Neural Information Processing Systems (NeurIPS), Workshop on Machine Learning in Structural Biology. -
Multimodal Few-Shot Learning with Frozen Language Models Neural Information Processing Systems (NeurIPS), Online. -
From Motor Control to Team Play in Simulated Humanoid Football doi:10.1126/scirobotics.abo0235 (Science Robotics), SR. -
Generative Art Using Neural Visual Grammars and Dual Encoders -
Game Plan: What AI can do for Football, and What Football can do for AI Journal of Artificial Intelligence (JAIR), ACM. -
Contrastive Training for Improved Out-of-Distribution Detection -
PolyGen: An Autoregressive Generative Model of 3D Meshes International Conference on Machine Learning (ICML), Vienna, Austria. -
Unsupervised Doodling and Painting with Improved SPIRAL -
A Hierarchical Probabilistic U-Net for Modeling Multi-Scale Ambiguities -
Data-Efficient Image Recognition with Contrastive Predictive Coding International Conference on Machine Learning (ICML), Vienna, Austria. -
Meta-Learning surrogate models for sequential decision making -
Attentive Neural Processes Neural Information Processing Systems (NeurIPS), Workshop on Bayesian Deep Learning, Montréal, Canada. -
Neural Processes International Conference on Machine Learning (ICML), Workshop on Theoretical Foundations and Applications of Deep Generative Models, Stockholm, Sweden. -
Conditional Neural Processes International Conference on Machine Learning (ICML), Stockholm, Sweden. -
Encoding Spatial Relations from Natural Language -
Learning models for visual 3D localization with implicit mapping Neural Information Processing Systems (NeurIPS), Workshop on Bayesian Deep Learning, Montréal, Canada. -
Consistent Jumpy Predictions for Videos and Scenes Neural Information Processing Systems (NeurIPS), Workshop on Bayesian Deep Learning, Montréal, Canada. -
A Probabilistic U-Net for Segmentation of Ambiguous Images Neural Information Processing Systems (NeurIPS), Montréal, Canada. -
Neural Scene Representation and Rendering Science, doi:10.1126/science.aar6170 -
Generative Temporal Models with Spatial Memory for Partially Observed Environments International Conference on Machine Learning (ICML), Stockholm, Sweden. -
Synthesizing Programs for Images using Reinforced Adversarial Learning International Conference on Machine Learning (ICML), Stockholm, Sweden. -
Machine Theory of Mind International Conference on Machine Learning (ICML), Stockholm, Sweden. -
Kickstarting Deep Reinforcement Learning -
Learning and Querying Fast Generative Models for Reinforcement Learning -
The Multi-Entity Variational Autoencoder Neural Information Processing Systems (NIPS), Workshop on Learning Disentangled Features, Long Beach, California, USA. -
Few-shot Autoregressive Density Estimation: Towards Learning to Learn Distributions International Conference on Learning Representations (ICLR), Vancouver, Canada. -
Emergence of Locomotion Behaviours in Rich Environments -
Unsupervised Learning of 3D Structure from Images Neural Information Processing Systems (NIPS), Barcelona, Spain. -
Attend, Infer, Repeat: Fast Scene Understanding with Generative Models Neural Information Processing Systems (NIPS), Barcelona, Spain. -
Reinforced Variational Inference Neural Information Processing Systems (NIPS), Workshop on Advances in Approximate Bayesian Inference, Montréal, Canada. Oral presentation. -
Kernel-Based Just-In-Time Learning for Passing Expectation Propagation Messages Uncertainty in Artificial Intelligence (UAI), Amsterdam, The Netherlands. -
Just-In-Time Kernel Regression for Expectation Propagation International Conference on Machine Learning (ICML), Workshop on Large Scale Kernel Learning, Lille, France. -
Consensus Message Passing for Layered Graphical Models International Conference on Artificial Intelligence and Statistics (AISTATS), San Diego, California, USA. -
Just-In-Time Learning for Fast and Flexible Inference Neural Information Processing Systems (NIPS), Montreal, Quebec, Canada. -
The PASCAL Visual Object Classes Challenge — a Retrospective International Journal of Computer Vision (IJCV), Springer. -
Assessing the Significance of Performance Differences on the PASCAL VOC Challenges via Bootstrapping Technical note. -
The Shape Boltzmann Machine: a Strong Model of Object Shape International Journal of Computer Vision (IJCV), Springer. CVPR special issue. -
A Generative Model for Parts-based Object Segmentation Neural Information Processing Systems (NIPS), Lake Tahoe, California, USA. -
The Shape Boltzmann Machine: a Strong Model of Object Shape Computer Vision and Pattern Recognition (CVPR), Rhode Island, USA. Oral presentation. -
Factored Shapes and Appearances for Parts-based Object Understanding British Machine Vision Conference (BMVC), Dundee, UK. Oral presentation.
Theses
-
Generative Probabilistic Models for Object Segmentation The University of Edinburgh, PhD thesis. -
Evolving Robust Control Strategies for Simulated Animats The University of Edinburgh, BEng dissertation.
Selected talks
-
AI in the Era of Large Language Models - TANK magazine (Apr 2023)
-
Neural Priors, Neural Encoders, Neural Decoders - Royal Society Workshop on New Approaches to 3D Vision (Nov 2021)
-
Representation Learning Without Labels - ICML 2020 tutorial (July 2020)
-
Making Machines Creative - TEDx Warwick (February 2020)
-
Neural Scene Representation and Rendering - Eurographics Symposium on Rendering, Strasbourg, France (July 2019)
- CVPR Workshop on Generative Models for 3D Understanding, Long Beach (June 2019)
- Sharif University of Technology, Tehran, Iran (December 2018)
- ICVGIP, Hyderabad, India (December 2018)
- The post binary conference, Frankfurt, Germany (November 2018)
- Re:work conference, London, UK (October 2018)
- Research@Google, Mountain View, California, USA (August 2018)
-
Attend, Infer, Repeat: Fast Scene Understanding with Generative Models - Beyond Supervised Learning Workshop, ICCV, Venice, Italy (October 2017)
- Sharif University of Technology, Tehran, Iran (December 2016)
- University of Tehran, Tehran, Iran (December 2016)
- Data Learning and Inference (DALI) Meeting, RNN Workshop, Sestri Levante, Italy (March 2016)
- Robotics Research Seminar, Oxford Universty, Oxford, UK (November 2016)
-
Modern Artificial Intelligence - Invited Lecture, Digikala, Tehran, Iran (December 2016)
- Machine Learning Tutorial Series, Imperial College London, London, UK (October 2016)
- Sharif University of Technology, Tehran, Iran (December 2015)
- Shahid Beheshti University, Tehran, Iran (December 2015)
- Tehran University, Tehran, Iran (December 2015)
- Campus London, London, UK (August 2015)
- Robotic's Society Lecture Series, King's College London, London, UK (March 2015)
- COMPM041 Guest Lecture, University College London, London, UK (March 2015)
-
Artificial Intelligence and Computer Aided Design - Alan Turing Institute, London, UK (November 2019)
- ETH Zurich, Zurich, Switzerland (July 2019)
- Foster + Partners Architects, London, UK (July 2019)
- Cartwright Picard Architects, London, UK (January 2019)
- Grimshaw Architects, London, UK (January 2015)
- Colloquium Seminar Series, University College London, London, UK (January 2015)
- Sharif University of Technology, Kish, Iran (December 2014)
- Faculty of Fine Arts, University of Tehran, Tehran, Iran (December 2014)
-
Extensions to Message Passing Inference - Sharif University of Technology, Kish, Iran (December 2014)
- Microsoft Research, Cambridge, UK (September 2014)
-
Boltzmann Machines and their Extensions - Heriot-Watt University, Edinburgh, UK (March 2013)
-
Generative Models of Images of Objects - Sharif University of Technology, Tehran, Iran (January 2013)
- Redwood Institute at the University of California at Berkeley, USA (December 2012)
- Bosch Research and Technology North America, Palo Alto, USA (December 2012)
- Toyota Technological Institute, University of Chicago, Chicago, USA (June 2012)
- New York University, New York, USA (June 2012)
-
The Shape Boltzmann Machine: a Strong Model of Object Shape - Computer Vision and Pattern Recognition (CVPR), Providence, USA (June 2012)
- Rank Prize Symposium on Machine Learning and Vision, Windermere, UK (March 2012)
-
Factored Shapes and Appearances for Parts-based Object Understanding - British Machine Vision Conference, Dundee, UK (August 2011)
- ACM Seminar, University of Tehran, Tehran, Iran (September 2011)
-
Hierarchical Probabilistic Models for Object Segmentation - Canadian Institute for Advanced Research Summer School, Toronto, Canada (August 2010)
-
High Dynamic Range Lighting on the Xbox 360 - TechMeetup, Edinburgh, UK (April 2010)
Other things
-
6 Things Everyone Should Know About Visualizing Research Results -
15 Jan 2015
- Guide
-
The Zeroth Rule of Presentations -
14 Oct 2012
- Guide
-
Patterns for Research in Machine Learning -
18 Jul 2012
- Guide
-
Thoughts on the Role of Supervision in Computer Vision -
18 May 2011
- Slides
Service
I co-organised the NeurIPS 2020 workshop on Muslims in Machine Learning, that was held virtually on December 8, 2020.
I co-organised the NIPS 2017 workshop on Machine Learning for Creativity and Design, that was held in Long Beach, California on December 8, 2017.
I co-organised the NIPS 2015 workshop on Black Box Learning and Inference, that was held in Montréal, Canada on December 12, 2015.
I regularly AC and review for NeurIPS, ICML, ICLR.
