Representation Extraction & Visualization

This exercise focuses on extracting and visualizing neural network representations using the MNIST dataset.

Key steps include:

• Training a small MLP with three hidden layers.
• Extracting hidden layer outputs.
• Projecting those representations into 2D using UMAP or T-SNE.
• Monitoring how the learned representations evolve over training.

The project provides insight into the learning dynamics of neural networks and highlights generalization by comparing train/test projection differences.