Debugging Neural Networks

Numerical experiments for better understanding neural networks.

Goal: Bridge the gap between practical behavior and theoretical models - small autoencoder example + detailed logging.

Installation: pip install -r requirements.txt

Configuration:

• Modify experiments/config.yaml according to your needs (input_dim, hidden_dims, lr, epochs, etc.).

Execution: python src/train.py

Results: A timestamped subdirectory will be created in the runs/ folder containing step_000000.npz files. These files include:

• param__ : weights / biases
• grad__ : gradients
• act__ : activations
• loss : loss at the time of saving

Analysis: python src/analysis.py # edit it, specify the run_dir and parameter names