Challenge

This project implements a Variational Autoencoder (VAE), a type of generative deep learning model, to learn compressed representations of fashion item images from the FashionMNIST dataset.

VAEs are powerful for unsupervised learning and generative tasks, making them valuable tools for dimensionality reduction and data generation in computer vision.

The main challenge was designing and training a VAE that could effectively encode and decode fashion images while maintaining reconstruction quality and exploring the structure of the learned latent space.

Implemented the encoder and decoder using PyTorch with reparameterization trick, KL divergence loss, and binary cross-entropy.

Analyzed the quality of image reconstruction and plotted 2D latent space to interpret learned representations