Somjit Nath

I am a PhD Student at McGill University/ Mila where I am advised by Professor Derek Nowrouzezahrai and Professor Samira Ebrahimi Kahou. I am currently working on representation learning for reinforcement learning. Previously, I was a Researcher at TCS Research, working under the supervision of Harshad Khadilkar, where I worked on applying Reinforcement Learning to Supply Chains and addressing the challenges that come with it.

I did my Masters from the University of Alberta where I was advised by Martha White. My thesis formulates a new approach for training Recurrent Neural Networks. I also worked on non-linear value function approximation for Reinforcement Learning. My undergrad was in Electrical Engineering from Jadavpur University, during which I worked mostly in the Internet of Things (IOT) domain along with medical imaging.

Email  /  CV  /  Twitter  /  Google Scholar  /  LinkedIn  /  Github

profile photo
Latest News
  • Our work on Spectral Temporal Contrastive Learning has been accepted at NeurIPS 2023 Workshop: Self-Supervised Learning - Theory and Practice.
  • Our work on Prioritizing Samples in Reinforcement Learning with Reducible Loss has been accepted to NeurIPS 2023.
  • Excited to announce I joined Borealis AI as a Machine Learning Research Intern for Summer 2023.
  • Our work on Learning object-centric representations with General Value Functions has been accepted at ICML 2023.
  • Our paper on Exploration with General Value Functions has been accepted at AAMAS Main Track 2023.
  • Locally Constrained Representations in Reinforcement Learning (joint work with Prof. Samira Ebrahimi Kahou) accepted to Deep RL Workshop, Neurips 2022.
  • Prioritizing Samples in Reinforcement Learning with Reducible Loss (joint work with Shivakanth Sujit, Pedro Braga and Prof. Samira Ebrahimi Kahou) accepted to Deep RL Workshop, Neurips 2022.
  • Our work on Handling Uncertain Lead Times in Supply Chains has been accepted at the 15th European Workshop on Reinforcement Learning (EWRL 2022).
Research

My main interests lie in the field of reinforcement learning. My fascination for RL developed after taking Rich Sutton's course during my Masters. In the context of RL, I am also interested in representation learning, model based RL and exploration.

3DSP Prioritizing Samples in Reinforcement Learning with Reducible Loss
Shivakanth Sujit, Somjit Nath, Pedro H. M. Braga, Samira Ebrahimi Kahou
Neural Information Processing Systems (NeurIPS), 2023
project / paper / code / slides
3DSP Discovering Object-Centric Generalized Value Functions From Pixels
Somjit Nath, Gopeshh Raaj Subbaraj, Khimya Khetarpal, Samira Ebrahimi Kahou
International Conference on Machine Learning (ICML), 2023
paper / code
3DSP Follow your Nose: Using General Value Functions for Directed Exploration in Reinforcement Learning
Durgesh Kalwar, Omkar Shelke, Somjit Nath, Hardik Meisheri, Harshad Khadilkar
International Conference on Autonomous Agents and Multiagent Systems (AAMAS), 2023
paper / poster
3DSP A Learning Based Framework for Handling Uncertain Lead Times in Multi-Product Inventory Management
Hardik Meisheri, Somjit Nath, Mayank Baranwal, Harshad Khadilkar
European Workshop on Reinforcement Learning (EWRL), 2022
paper
3DSP Revisiting State Augmentation methods for Reinforcement Learning with Stochastic Delays
Somjit Nath, Mayank Baranwal, Harshad Khadilkar
Conference on Information and Knowledge Management (CIKM), 2021
paper / code / slides / poster
3DSP Scalable multi-product inventory control with lead time constraints using reinforcement learning
Hardik Meisheri, Nazneen N Sultana, Mayank Baranwal, Vinita Baniwal, Somjit Nath, Satyam Verma, Balaraman Ravindran, Harshad Khadilkar
Neural Computing and Applications (NCA) , 2021.
paper
3DSP SIBRE: Self Improvement Based REwards for Adaptive Feedback in Reinforcement Learning
Somjit Nath, Richa Verma, Abhik Ray, Harshad Khadilkar
International Conference on Autonomous Agents and Multiagent Systems (AAMAS), 2021.
paper
3DSP Training Recurrent Neural Networks Online by Learning Explicit State Variables
Somjit Nath, Vincent Liu, Alan Chan, Xin Li, Adam White, Martha White
International Conference on Learning Representations(ICLR), 2020.
paper / slides / code
3DSP Two-Timescale Networks for Nonlinear Value Function Approximation
Wesley Chung, Somjit Nath, Ajin Joseph, Martha White
International Conference on Learning Representations(ICLR), 2019.
paper
3DSP A Fixed-Point Formulation for Recurrent Neural Networks
Somjit Nath, Taher Jafferjee and Martha White
Continual Learning Workshop, Neural Information Processing Systems(NeurIPS) , 2018.
paper
3DSP Rejection Sampling for Off-Policy Learning
Wesley Chung, Sina Ghiassian, Somjit Nath, and Martha White
Continual Learning Workshop, Neural Information Processing Systems(NeurIPS) , 2018.
paper
3DSP Smartphone Camera Based Analysis of ELISA using Artificial Neural Network
Somjit Nath, Subhannita Sarcar, Biswendu Chatterjee, Nabendu Chatterjee, Rhishita Chourashi
IET Computer Vision , 2018.
paper
3DSP Arduino Based Door Unlocking System with Real Time Control
Somjit Nath, Paramita Banerjee, Rathindra Nath Biswas, Swarup Kumar Mitra and Mrinal Kanti Naskar
International Conference on Contemporary Computing and Informatics (IC3I) , 2016.
paper

Source code and style from Jon Barron's website.