@anirudh-p1

This project was made for the Non-Trivial Fellowship. It is a machine learning paper testing various attention kernels and their effects on mitigating superposition, and whether this depends on compression size.

Anti-MABL stands for Axial Neutrality Training Instrument and is an ML-powered, contactless electromagnetic exercise countermeasure designed to mitigate (upper-body) skeletal degradation and bone mass loss in microgravity. It uses a frictionless eddy current brake combined with a tension load cell and closed-loop control to provide adaptive resistance.

The Idea was to create an alternative complete attention mechanism that could act as a drop in replacement for softmax. It replaces stochastic, unconstrained heuristic logits with deterministic kernel weightings derived from Ramanujan's third-order mock theta functions and classical q -series and introduces an approximate modular-symmetry inductive bias directly into the forward pass of Transformer architectures.

Build and train a model on HRV to get it to give predictions on fatigue/recovery like Whoop does, but better.

RF Jamming Detection via Complex-Valued Spectogram Autoencoding

Physics-Informed Neural Networks to predict Metabolic Thresholds