Projects

Accelerating molecular sampling (report) Using generative models to speed up parallel tempering by learning transports between Boltzmann distributions.

RNA design models (preprint) Improving RNA inverse folding with greater expressivity and biologically inspired priors (an enhanced gRNAde).

Memory in recurrent networks (report) Exploring how “quenched variability” in neural connectivity can expand memory storage capacity.

Few-shot relation classification Combining DistilBERT with CAVIA for efficient few-shot learning on relation classification tasks.

Reinforcement learning in grid-worlds Implementing and comparing dynamic programming and model-free RL algorithms across classic grid-world scenarios.

Bayesian inference in high dimensions Exploring scalable variational methods and dimensionality reduction to tackle inference in complex, high-dimensional models.

Speech recognition with CTC Building an end-to-end ASR system on the TIMIT dataset using CTC to achieve competitive error rates with improved alignment.

Glomeruli tissue segmentation Training a U-Net to segment kidney glomeruli functional units for precise biomedical image analysis (PDF).

Cross-domain sentiment classification Applying adversarial domain adaptation to make sentiment classifiers robust across diverse domains.

Amari's neural field simulations Modeling spatial firing patterns with Amari's neural field equations to study stability and cortical dynamics.

Power-seeking AI analysis A theoretical essay on the emergence of power-seeking behavior in advanced AI and its implications for alignment.