Andrew (Drew) Jaegle

I'm a Research Scientist at DeepMind in London, where I work on artificial intelligence with an amazing group of collaborators.

Before DeepMind, I worked on computer vision and computational neuroscience at the University of Pennsylvania, where I did a PhD in the GRASP Lab with Kostas Daniilidis and a postdoc with Nicole Rust. I also spent time in the labs of Diego Contreras and Maria Geffen and with the folks in the Computational Neuroscience Initiative. During my PhD, I did research internships at the Max-Planck Institute for Intelligent Systems, where I worked with Michael Black and Javier Romero, and at DeepMind, where I worked with Greg Wayne. Prior to that, I worked on consciousness and attention at the City University of New York with Tony Ro. I have a BA from Texas A&M University, where I studied philosophy, music composition, and mathematics.

Email  /  CV  /  Google Scholar  /  Twitter  /  LinkedIn

Research

My research is focused on topics in AI including vision and sensory reasoning, imitation, and unsupervised learning. I'm broadly interested in perception and behavior, and I also like to think about and work on philosophy and music.

Perceiver IO: A General Architecture for Structured Inputs & Outputs

A Perceiver architecture that handles arbitrary inputs and outputs: SoTA on Sintel optical flow, matches a strong BERT baseline but without tokenization, simultaneous multi-task GLUE fine-tuning, better/faster results on ImageNet without 2D assumptions, and strong results on StarCraft II & multimodal Kinetics autoencoding.

Which priors matter? Benchmarking models for learning latent dynamics

A new benchmark for models of physical dynamics from images with 17 datasets and strong baselines. Code and data coming soon!

Imitation by Predicting Observations

Imitation from observations by sequence modeling using a framework for robust, non-adversarial inverse reinforcement learning.

Perceiver: General Perception with Iterative Attention

A scalable, domain-agnostic attentional architecture that works on images, audio, point clouds, video, multimodal data and more.

Temporal Difference and Return Optimism in Cooperative Multi-Agent Reinforcement Learning

How to use optimism to encourage agents to learn in the presence of other, suboptimal agents in multi-agent reinforcement learning.

Game Plan: What AI can do for Football, and What Football can do for AI

An overview of the challenges in football analytics and how tools from AI (including computer vision, game theory, and statistical learning) can give new and better insights.

Physically embedded planning problems: New challenges for reinforcement learning

A suite of difficult Mujoco domains that pair high-level, abstract structure with physical embodiment. These domains are very challenging for current black-box RL methods, but can be solved by building in high-level structure: we challenge you to bridge this gap!

Beyond Tabula-Rasa: a Modular Reinforcement Learning Approach for Physically Embedded 3D Sokoban

We revisit the classic sense-plan-act structure to build a modular reinforcement learning agent that can efficiently solve physical tasks with abstract structure.

Hamiltonian generative networks

Methods for learning the structure of physical systems from video without supervision using generative models built with ideas from Hamiltonian mechanics.

Keyframing the Future: Keyframe Discovery for Visual Prediction and Planning

A model that first predicts important moments in the future and then uses these to infer the missing details, without supervision.

Visual novelty, curiosity, and intrinsic reward in machine learning and the brain

Connections between two literatures with important, but largely unappreciated, connections: intrinsic motivation in AI and visual novelty in neuroscience and psychology.

Population response magnitude variation in inferotemporal cortex predicts image memorability

A simple signature of an image’s memorability – how easy it is for people to remember, compared to other images – that is shared by the brain and convolutional neural networks.

Codes, functions, and causes: A critique of Brette's conceptual analysis of coding

An argument for the validity of functional decomposition and approximation in causal analyses of complex systems like brains and deep networks (also tents). A response to Brette 2018 (and see his response to our response).

Learning what you can do before doing anything

An unsupervised method that captures the latent structure of behavior from video using compositional video prediction models.

Predicting the Future with Transformational States

Methods for video prediction by reasoning about changes to static content in a learned latent space. Check out Oleh’s blog post discussing some of the problems with using standard reconstruction losses.

Understanding image motion with group representations

A simple method for unsupervised learning of motion in videos using ideas from group theory.

Fast, robust, continuous monocular egomotion computation

Robust methods for estimating camera motion in real-world video by reasoning about the contribution of each element of an optic flow field to the global motion estimate.

Emergence of invariant representation of vocalizations in the auditory cortex

How nuisance transformations – changes to a sound that don’t change its meaning, like saying the same word faster or slower – are tuned out as signals moves from an earlier to a later stage of auditory cortex.

Second order receptive field properties of simple and complex cells support a new standard model

A critical review of Fournier et al’s work characterizing the structure of synaptic receptive fields in visual cortex.

Direct control of visual perception with phase-specific modulation of posterior parietal cortex

Evidence for a causal role of the phase of alpha-band (10 Hz) oscillations in visual perception.

This guy does good work.