Variational Autoencoder on FPGA

A 16-bit fixed-point quantized VAE for processing MNIST on a Nexys 4 FPGA with near-instantaneous inference.

Course Project – Mentor: Prof. Joycee Mekie, IIT Gandhinagar

  • Designed and implemented a fixed-point quantized Variational Autoencoder for processing the MNIST dataset on an FPGA.
  • Verified image reconstruction fidelity against software-based PyTorch models.
  • Achieved latency reductions enabling near-instantaneous inference.