Somjit Nath

PhD Student, Electrical and Computer Engineering, McGill University and Mila, Montréal

I am a PhD student at McGill University and Mila, advised by Derek Nowrouzezahrai and Samira Ebrahimi Kahou. My research focuses on reinforcement learning, representation learning, model based RL, and real world sequential decision making.

I am interested in how agents can learn structured, predictive representations of their environment to enable efficient learning, long horizon reasoning, and generalization in complex environments.

Recently, I was a Research Intern at Microsoft Research Cambridge, where I worked on improving sample efficient imitation learning in video based environments. I was also a Scientist in Residence at 4Division and Mila, where I worked on scalable Robotics Transformer models for real world robotic learning.

Previously, I was a Machine Learning Research Intern at RBC Borealis AI and a Researcher at TCS Research, where I worked on reinforcement learning for continuous time event data and supply chain optimization.

Email CV Google Scholar LinkedIn GitHub Twitter
Somjit Nath profile photo

Latest News

  • Joined Microsoft Research Cambridge as a Research Intern working on sample efficient imitation learning.
  • Served as Scientist in Residence at 4Division and Mila working on Robotics Transformer models.
  • Our paper Behaviour Discovery and Attribution for Explainable Reinforcement Learning appeared in TMLR 2025.
  • Our paper Unsupervised Event Outlier Detection in Continuous Time was accepted at the NeurIPS 2024 Self Supervised Learning Workshop.
  • Our work on Task Oriented Slot Based Cumulant Discovery in General Value Functions was selected as a spotlight at the RLC 2024 workshop.
  • Received Outstanding Reviewer Awards at ICML 2025 and ICCV 2023.

Research

My research focuses on reinforcement learning, representation learning, and model based decision making. I am particularly interested in how agents can learn predictive and structured representations that support efficient learning, exploration, planning, and generalization in complex environments.

More broadly, I am interested in bridging reinforcement learning with world models, self supervised learning, imitation learning, and real world sequential decision problems.

Experience

Research Intern
May 2025 – Aug 2025
Microsoft Research Cambridge, United Kingdom
Worked on improving representation learning approaches for sample efficient imitation learning in video game environments.
Scientist in Residence
Jun 2024 – Sep 2024
4Division and Mila, Montréal, Canada
Worked on scalable Robotics Transformer models trained on diverse robotic data for strong real world generalization.
Machine Learning Research Intern
May 2023 – Aug 2023
RBC Borealis, Montréal, Canada
Developed an unsupervised outlier detection framework in continuous time event sequences using reinforcement learning.
Researcher, Data and Decision Sciences
Nov 2019 – Jan 2022
Tata Consultancy Services Research and Innovation, Mumbai, India
Applied reinforcement learning to multi product, multi node inventory management, improving performance over existing industry approaches, and developed RL methods for delayed actions and observations.
Research Intern
May 2015 – Jul 2015
Indian Statistical Institute, Kolkata, India
Worked on OCR and character segmentation methods for Bengali and English text.

Selected Publications

Behaviour Discovery and Attribution for Explainable Reinforcement Learning
Rishav Rishav, Somjit Nath, Vincent Michalski, Samira Ebrahimi Kahou · TMLR 2025
Paper
Unsupervised Event Outlier Detection in Continuous Time
Somjit Nath, Kry Yik-Chau Lui, Siqi Liu · NeurIPS Workshop 2024
Paper
Task Oriented Slot Based Cumulant Discovery in General Value Functions
Vincent Michalski, Somjit Nath, Derek Nowrouzezahrai, Doina Precup, Samira Ebrahimi Kahou · RLC Workshop 2024 (Spotlight)
Paper
Spectral Temporal Contrastive Learning
Sacha Morin, Somjit Nath, Samira Ebrahimi Kahou, Guy Wolf · NeurIPS Workshop 2023
Paper
Prioritizing Samples in Reinforcement Learning with Reducible Loss
Shivakanth Sujit, Somjit Nath, Pedro H. M. Braga, Samira Ebrahimi Kahou · NeurIPS 2023
Project Paper Code
Discovering Object-Centric Generalized Value Functions From Pixels
Somjit Nath, Gopeshh Raaj Subbaraj, Khimya Khetarpal, Samira Ebrahimi Kahou · ICML 2023
Paper Code
Follow your Nose: Using General Value Functions for Directed Exploration in Reinforcement Learning
Durgesh Kalwar, Omkar Shelke, Somjit Nath, Hardik Meisheri, Harshad Khadilkar · AAMAS 2023
Paper
Revisiting State Augmentation Methods for Reinforcement Learning with Stochastic Delays
Somjit Nath, Mayank Baranwal, Harshad Khadilkar · CIKM 2021
Paper Code
SIBRE: Self Improvement Based Rewards for Adaptive Feedback in Reinforcement Learning
Somjit Nath, Richa Verma, Abhik Ray, Harshad Khadilkar · AAMAS 2021
Paper
Training Recurrent Neural Networks Online by Learning Explicit State Variables
Somjit Nath, Vincent Liu, Alan Chan, Xin Li, Adam White, Martha White · ICLR 2020
Paper Code

Awards and Service

  • Outstanding Reviewer, ICML 2025
  • Outstanding Reviewer, ICCV 2023
  • McGill Engineering Doctoral Award
  • TCS Citation Award, 2021 and 2022
  • Scholarship for Academic Excellence, State Electrical Engineers’ Association

Education

  • McGill University and Mila
    PhD in Electrical and Computer Engineering, 2022 – present
  • University of Alberta
    MSc in Computing Science, 2017 – 2019
  • Jadavpur University
    BE in Electrical Engineering, 2013 – 2017