Learning more about manifold learning algorithms has been on my TODO list for a while. I also wanted to play with Captum for a while as well. The feature_analysis project started with the desire to dive into the details of t-SNE and UMAP and the willingness to dive further into the neural network feature, specifically convolutions, and then onto feature-based searches.

In the current state, this project completes the deep dive into UMAP and t-SNE and applies that to multiple datasets. My detailed notes on this topic are documented in this [bog] titled Review and comparison of two manifold learning algorithms: t-SNE and UMAP.

Early-stage exploration of Captum and weights and feature analysis can be found in this notebook. However, in many ways, this is a work in progress.

This project will be a ongoing slow burn person project for me to exlore into [activation atlas] and Captum and build on feature based indexers.