Toy Ideas
Preliminary ideas spanning theoretical machine learning, interpretability, and geometry-aware learning.
Do Neural Networks Memorize Without Separation?
This project explores how neural networks memorize noisy labels through representation analysis. The results show that memorization does not require separating noisy samples into distinct clusters. Instead, noisy and clean data remain geometrically entangled, while the model fits noise using high-confidence decision boundaries within the same representation space.
Decodability Is Not Sufficiency
A mechanistic interpretability study showing that highly decodable and ablation-sensitive features are not necessarily sufficient for correct model behavior. The project uses controlled counterfactual evaluation to distinguish sensitivity from sufficiency in representation-level analysis.
Heat Kernel SVM on Manifolds
A geometry-aware classification project using support vector machines with heat kernels defined on manifolds. The implementation studies both a flat torus with an analytic Fourier heat kernel and a general surface setting using a mesh Laplacian and spectral heat kernel construction.