The Hopfield model is a fully connected neural network of biological inspiration that aims to reproduce an associative memory. The most important contribution on the topic is the phase diagram obtained via the replica trick, which makes this system one of the few analytically treatable model of neural network.
The scope of an associative memory is somehow different from that of most deep learning networks. Despite this, we can observe some phenomenological similarities with modern machine learning concepts. This might suggest that this system could be of key interest for a general theory of neural computation.
During the seminar we will also discuss some recent generalizations of the model that narrow the gap with the deep neural networks we are used to. Specifically, we will talk about the use of structured data and we will take a look at the DayDreaming algorithm that can improve the network capacity up to its theoretical limit.