This repository contains experimentation on dense associative memory modules, based on the theory proposed by Krotov and Hopfield.
Traditional Hopfield network, Dense Associative network and spherical memory have been implemented incontext of image retension and retrieval. The goal here is to analyze how good are these memory modules at learning complex patterns.
All three of these networks were able to learn binary images easily with perfect recall during inference.
Only dense associative network and spherical memory is tested on 64x64 RGB images.
Both architectures fail to retrieve the proper image during the inference. This points towards the possibility that the memory landspace has collapsed into a single metastable state. This would cause the model to output a mean of all the stored patterns no matter what the input pattern is. This collapse happens even when we only save 2 images and only 10% of the image is perturped.
