Memory Game

Compete against an AI model that is trained on-device in seeing who can best guess what the noisy image is depicting.

Under the hood, my custom implementation of the Sparse Quantized Hopfield Network proposed by Alonso and Krichmar written in Rust, uses Hebbian learning rules to update its parameters to learn a robust retrieval mechanism resistant to noisy inputs. By eschewing backpropagation and utilizing sparse activations, this Hopfield network is able to memorize images after seeing them only once, making it suitable for on-device training and other compute constrained settings.

My original PyTorch implementation can be found in github, along with the Rust translation, and the game itself.

Original paper:

Alonso, N., & Krichmar, J. (2023). A Sparse Quantized Hopfield Network for Online-Continual Memory. https://arxiv.org/abs/2307.15040