Portfolio

Machine Unlearning Permalink

Implemented common machine unlearning (MU) algorithms such as zero-glance and zero-shot unlearning via error-maximizing noise and gated knowledge transfer respectively, selective synaptic dampening (SSD), incompetent teacher unlearning, etc., exploring the nascent field of MU and its applications in the right to be forgotten, debiasing, influence functions and model interpretability, and more.

TinyTransformer Permalink

Built and trained decoder transformer models in parallel from scratch on TinyStories, investigated scaling laws of dataset and model size with validation loss and created story generation demo.

Art Generation with GANs Permalink

Implemented and compared DCGANs and Creative Adversarial Network (CAN)s to generate paintings, performed hyperparameter tuning and metric evaluations, developed interactive Streamlit demo.

Mixture-of-Experts Implementation Permalink

Implemented a Switch Transformer alongside a conventional autoregressive transformer and trained on TinyShakespeare to research effects of mixture-of-experts architecture on validation loss, sample-efficiency and training time.