Currently, I am a Research Scientist at Oracle Labs working on coding agents, following a DPhil (PhD) at the University of Oxford. During my DPhil in the WhiRL lab, I was supervised by Shimon Whiteson, funded by the Oxford-Google DeepMind Doctoral Scholarship, and studied deep reinforcement learning (RL). My main research area was meta-RL. (For an introduction, see this video tutorial, this panel, or this podcast interview where I explain meta-RL!) I've worked on hypernetworks, initialization methods, and sequence models for in-context learning – in addition to authoring a survey of meta-RL. Previously, I did my MS and BS at Brown University, completed a pre-doc at Microsoft Research on sequence models in RL, and researched autonomous vehicles in both academia and industry.
At Brown University, I was advised by Michael Littman. My research focused on human feedback, imitation learning, and multi-agent game theory. Some of our work in the self-driving car lab gained publicity in New Scientist. Other projects included: an RL agent in Minecraft using emotion detection as feedback and a GAN to reconstruct corrupted images. As a TA for the first iteration of Brown's graduate-level deep learning course, I designed a lab and gave a guest lecture on sequence-to-sequence machine translation.
In industry, I worked at Microsoft, Lyft, Adobe, and several smaller companies. I completed a pre-doc at Microsoft Research with Katja Hofmann on long-term memory in RL. At DeepScale, acquired by Tesla, I worked on perception for autonomous vehicles, developing novel methods for instance segmentation. At Lyft I designed a framework for sequential decision-making problems, including a special-case solver specific to autonomous vehicles at stop intersections. At Adobe I built neural networks to forecast marketing data. I also worked at a robotics startup on software and hardware, co-created Food with Friends (an iOS app), and researched in-context learning with large language models (LLMs) for proteins at InstaDeep.
Broadly, my interests include generalization, adaptation, and representation. Specifically, topics include learning to learn (in-context), sequence models, and few-shot learning – including intersections with multi-agent RL and long-horizon learning. I am also interested in human (and AI) feedback, vision-language models for environment design, hypernetworks, and any challenging problem in ML. See my CV, Google Scholar, and GitHub for past work.
For mentorship and research supervision, you can sign up for my office hours through the ML Collective (MLC). To discuss joint research supervision for students or researchers in your group, feel free to reach out by email.
JakeABeck [at] gmail [dot] com
Jake.Beck [at] oracle [dot] com