Research
Featured Projects
Accelerating Molecular Dynamics with Deep Generative Models
Designing neural samplers to accelerate molecular simulation by learning efficient approximations to the Boltzmann distribution. Leveraging diffusion models and energy-based methods to improve sampling of equilibrium configurations.
In progress →gRNAdeX: eXpressive, Biologically-eXtensible gRNAde
Improved the gRNAde model for RNA inverse folding by refining geometric expressivity, pooling mechanism, sampling robustness, and incorporating biologically inspired priors.
View project →Few-Shot Relation Classification with DistilBERT & CAVIA
Few-shot learning for relation classification on the FewRel dataset by combining a frozen DistilBERT encoder with the CAVIA meta-learning framework. Enables fast N-way K-shot adaptation via task-specific context vectors and includes configurable training and evaluation scripts.
View project →Reinforcement Learning in Grid-World Environments
Implemented a comprehensive suite of RL algorithms—ranging from dynamic programming (Policy Iteration, Value Iteration) to model-free methods (SARSA, Q-learning, Expected SARSA)—across multiple grid-world scenarios.
View project →Bayesian Inference in High Dimensions
Bayesian inference in high-dimensional parameter spaces, combining variational approximation techniques with dimension reduction approaches to address scalability challenges in complex statistical models.
View project →TIMIT Speech Recognition Using CTC
Implemented an end-to-end automatic speech recognition system using Connectionist Temporal Classification (CTC) on the TIMIT speech corpus, achieving competitive phone error rates through architectural innovations in the acoustic model and improved alignment strategies.
View project →Other Research Projects
Enhancing Memory Storage in Neuronal Networks
Investigated how introducing "quenched variability" in the connectivity of recurrent neuronal networks can improve their capacity to store and recall spatial memories.
Glomeruli Tissue Unit Segmentation
A U-Net-based segmentation process for precise identification and analysis of Glomeruli Functional Tissue Units in kidney tissues using advanced computer vision techniques.
Cross-Domain Sentiment Classification
An Adversarial Domain Adaptation (ADA) approach for Cross-Domain Sentiment Classification (CDSC) that enhances model robustness across diverse domains.
Amari's Neural Field Simulations
Research on Amari's equations to simulate spatial firing patterns across neural fields, exploring their stability and dynamics in depicting cortical activity.
Power-Seeking AI Analysis
Essay on the emergence of power-seeking behavior in advanced AI systems and its implications for AI safety and alignment. Includes formal models and theoretical frameworks.