/Disentanglement-in-discriminative-models

Improve generalizability of CNN-based vision models to downstream tasks by disentangling learned representation space dimensions

Primary LanguagePythonMIT LicenseMIT

Disentanglement-In-Discriminative-Models

Course project for CS 8803 LS

We improve generalizability of CNN-based vision models to downstream tasks by disentangling learned representation space dimensions.

Our motivation, approach and results can be found in this course presentation.

This repository contains the first known PyTorch (AutoGrad compatible) implementation of the KSG Mutual Information estimator.