Projects
Here are some of the projects I’ve worked on over the years, mostly as part of graduate courses or during my free time.
Avoiding Collapses in Non-Contrastive SSL Methods
Investigated how recent non-contrastive self-supervised learning (SSL) methods prevent collapse and maximize representation information without negative pairs. Focused on architectural changes and loss regularizations enabling high information content in representations.
Neural Decoding from Primary Motor Cortex
Behavior Decoding: Designed a vision system to track free-behaving rats implanted with microelectrode arrays. After recording sessions, neural signals were processed to extract spike trains and paired with corresponding behaviors. RNN-based neural decoders were then trained to classify these behaviors.
Force Decoding: Trained neural decoders to predict applied force on a knob sensor from rat motor cortex activity using different LSTM architectures and loss formulation.
Kinematics Decoding: Developed a pipeline to decode kinematic targets from intra-cortical neural recordings using deep recurrent networks. Led a team of research students in creating a codebase for signal processing, dataset generation (customizable by parameters such as temporal resolution and window length), and decoder training.
Inductor Magnetic Energy Estimation Based on Surrogate Models
Explored ML/DL models as fast surrogates to approximate mappings made by Finite-Element Analysis (FEA) for evaluating inductor designs. Used a proprietary dataset of 100,000 samples to predict magnetic energy based on design topology, reducing evaluation time drastically.
Simulation and Optimization of Wood Drying through Deep Reinforcement Learning
Developed a simulator using OpenAI Gym and Simpy to optimize wood drying processes with a Deep-RL agent. Employed PPO with heuristic-guided reinforcement learning to improve resource allocation and demonstrated potential for industrial process optimization.
Neural Combinatorial Solver using Reinforcement Learning
Compared end-to-end neural solvers with traditional methods for solving combinatorial optimization problems. Focused on Weapon-Target Assignment (WTA) using a graph attention network and reinforcement learning to learn optimal policies minimizing destructive target values.
Critical Temperature Prediction for Superconducting Materials
Used deep neural networks to predict critical temperatures of superconductors based on features like atomic weights and entropy. Leveraged a dataset of 21,263 materials to map material properties to laboratory-measured critical temperatures.
Trade-offs Between Model-Free and Model-Based Reinforcement Learning Methods
Explored trade-offs between model-free and model-based RL approaches, with a focus on implementing and understanding MuZero, highlighting its ability to combine planning and learning efficiently.
Theory Behind ML Models + Code Implementations
Created an 80-page document summarizing machine learning models with theoretical explanations and Python code implementations. Compiled from 20 books and various papers, it also served as a learning resource for friends.