// the find
SkalskiP/ILearnDeepLearning.py
This repository contains small projects related to Neural Networks and Deep Learning in general. Subjects are closely linekd with articles I publish on Medium. I encourage you both to read as well as to check how the code works in the action.
A grab-bag of small notebooks and scripts, each paired with a Medium post, walking through neural network fundamentals from scratch: backprop math in plain NumPy, optimizers, overfitting/regularization, a from-scratch CNN, and a comparison of explainability libraries (ELI5, LIME, SHAP). It's aimed at people who want to see the math behind the Keras API calls rather than at anyone looking for a reusable library.
The NumPy-only neural net and CNN implementations are genuinely useful for understanding backprop and convolution mechanics without a framework hiding the linear algebra. Each project is tied to a full-length article, so the code has real prose explaining the 'why' instead of being a bare script. The optimizer comparison and classification-boundary GIFs are unusually good pedagogical visuals — more informative than the usual loss-curve screenshot. The explainability shootout (ELI5 vs LIME vs SHAP on the same ResNet) is a decent reference for picking a tool, not just a demo of one.
This is a pile of independent notebooks, not a library — every subfolder has its own environment.yml or requirements.txt, so there's no single install path and the top-level 'Hit the ground running' instructions don't actually get you running anything specific. Last push is December 2023 but most content is Keras/TF-1-era circa 2018-2019 (there's even a Python 3.5 .pyc committed), so expect dependency rot on any current Python/TensorFlow setup. Only one of the ~10 sub-projects (the CNN one) has actual unit tests; the rest are notebooks with no verification beyond 'it ran for me once.' No CI, so there's no signal on whether any of this still executes today.