A fully custom implementation of AutoEncoders and Variational AutoEncoders (VAE) built from the ground up in NumPy—no TensorFlow, no PyTorch. Features a modular architecture with encoder-decoder symmetry, stochastic sampling layers for VAEs, and support for multiple optimizers (SGD, Momentum, Adam) with comprehensive loss functions (MSE, MAE, Cross-Entropy).

Highlights

• Custom backpropagation engine with dropout regularization
• Font character reconstruction & latent space visualization
• Stochastic layer implementation for VAE reparameterization trick
• Modular activation functions (ReLU, Sigmoid, Tanh) via Strategy pattern
• Complete with unit tests and training metrics tracking

Features

• AutoEncoder Architecture: Encoder-decoder symmetry with configurable hidden layers
• Variational AutoEncoder (VAE): Full implementation with KL divergence loss and reparameterization trick
• Multiple Optimizers: SGD, Momentum, and Adam with adaptive learning rates
• Loss Functions: MSE, MAE, and Cross-Entropy for flexible training objectives
• Regularization: Dropout layers for preventing overfitting
• Visualization Tools: Latent space exploration and reconstruction quality analysis

Built for the Artificial Intelligence Systems course at Buenos Aires Institute of Technology (ITBA).