Misha Denil

Career

Misha Denil's homepage describes 10 years as a Research Scientist at Google DeepMind, followed by a current period of working independently. Denil completed a PhD in the Computer Science department at the University of Oxford supervised by Nando de Freitas, after a master's degree at the University of British Columbia, where Denil also worked with de Freitas.¹

Research

From 2019 to 2024 Denil worked on data-driven robotics, at the meeting point of large-scale machine learning and real robots. Denil groups this work into learning from demonstrations, reward learning, real-world evaluation and complete agent systems, with papers such as Task-Relevant Adversarial Imitation Learning (CoRL 2020), Positive-Unlabeled Reward Learning (CoRL 2021, with Danfei Xu) and RoboCat, a self-improving generalist agent for robotic manipulation.²³⁴

Earlier projects included learning to learn (2016–2017), among them the paper "Learning to learn by gradient descent by gradient descent," which treats the design of optimization algorithms as a learning problem, and learning to perform physics experiments with deep reinforcement learning (2016). Before that, Denil worked on predicting parameters in deep learning, deep learning for natural language processing, and the consistency of random forests.²³⁵

Work and contributions

• RoboCat: A Self-Improving Generalist Agent for Robotic Manipulation Published research

  Co-author of the 2023 RoboCat paper (Transactions on Machine Learning Research), a visual goal-conditioned decision transformer trained on experience from simulated and real robotic arms.⁴³

• Learning to learn by gradient descent by gradient descent Published research

  Co-author, listed with Google DeepMind, of the 2016 paper showing optimization algorithms implemented as LSTMs that learn to exploit problem structure.⁵³

• Predicting Parameters in Deep Learning Published research

  First author of the NeurIPS 2013 paper, written with Babak Shakibi, Laurent Dinh, Marc'Aurelio Ranzato and Nando de Freitas, on predicting parameters in deep networks.³²

Timeline

1. 2010Published "Overlap versus Imbalance," with Thomas Trappenberg, at the Canadian Conference on Artificial Intelligence.³
2. 2013Published "Consistency of Online Random Forests" at ICML and "Predicting Parameters in Deep Learning" at NeurIPS.³
3. 2014Published "Narrowing the Gap" on random forests in theory and practice at ICML 2014.³
4. Jun 14, 2016Co-authored "Learning to learn by gradient descent by gradient descent."⁵
5. 2017Co-authored "Learned Optimizers that Scale and Generalize" and the Programmable Agents technical report.³
6. 2019Began several years of work on data-driven robotics, running through 2024.²
7. Jun 20, 2023Co-authored RoboCat, a self-improving generalist agent for robotic manipulation.⁴
8. 2024Published "Rapid Object Annotation" as a Google DeepMind technical report.³

Documented roles

The sources do not confirm a current role; these are the roles they document.

• Research Scientist (10 years), Google DeepMind¹⁵

Companies in their history

• Google DeepMind 64 people in the wiki

Related people

• Tuomas HaarnojaConnected to Google DeepMind
• Justin FuConnected to Google DeepMind
• Tom SchaulConnected to Google DeepMind
• Oriol VinyalsConnected to Google DeepMind
• Richard SuttonConnected to Google DeepMind
• Gavriel StateRobotics · Reinforcement learning
• Deepak PathakRobotics · Reinforcement learning
• Vikash KumarRobotics · Reinforcement learning
• Pieter AbbeelRobotics · Reinforcement learning
• Jonathan HurstRobotics · Reinforcement learning