Address: Microsoft Research,
300 Lafayette Street,
New York, NY 10012
Dipendra Misra
I am a Senior Researcher at Microsoft Research, New York.
I received my PhD in computer science from Cornell University (2019) and my
bachelors in computer science from Indian Institute of Technology Kanpur (2013).
Research Interests:
My main interest is in developing efficient machine learning algorithms with
applications to real-world problems. The word efficient here includes provable,
sample and computationally efficient, interpretable, scalable, and ethical.
My empirical focus is on problems in natural language understanding and allied fields.
I am currently active in reinforcement learning theory, interactive learning,
representation learning, and language and vision problems. I also have interest
in computational social science, and data and society.
News: We have a new paper on interactive learning at ICML 2021 that only uses language feedback (i.e., no reward, actions, etc.).
Our algorithm uses language descriptions of trajectories and solves a sequence of supervised learning problems. We evaluate on
grounded language understanding tasks and provide convergence guarantees.
[arXiv paper]
News: Qinghua Liu and I recently gave a talk at RL theory seminar. You can find it
here.
Provable RL: We have three new provable reinforcement learning algorithms for rich-observation problems.
These algorithms are computationally efficient, and their sample complexity is either independent of the size of observation space, or
only weakly depends on it.
-
FactoRL at ICLR 2021 that provably solves a subset of rich-observation problems with a latent exponentially large state space.
FactoRL learns the latent factorized model, and a state decoding function.
[Paper] [Code coming soon] -
RichID at NeurIPS 2020 that provably solves continuous control problem
with latent LQR dynamics and rich-observations. RichID learns the latent LQR model and a near-optimal policy.
[arXiv Version] [Code] -
Homer at ICML 2020 that solves rich-observation problems with a discrete latent state space. Homer provably explores,
recovers the latent dynamics, and optimizes any given reward function.
[arXiv Version] [ICML Version] [Code] [Medium Post]
- For post-doc and full-time positions in reinforcement learning apply here.
Quick Links: MSR Reinforcement Learning, Cereb-RL Code Base, CIFF Code Base, My Blog, A Bandit Game, RL Formulas
