Bifurcations for mechanisms of memory formation in a neural network #

Adam Essex, Natalia Janson, Alexander Balanov

Poster session

Abstract #

Neural networks are abundant in applications across a variety of fields and their importance in advancing technology is becoming ever more clear. However, there is a fundamental lack of transparency in the mechanisms underpinning their intelligent behaviour even in the simplest systems, namely, those responsible for memory formation and decision making. Many of the attempts to combat this issue have so far been unsuccessful and hence we propose a different approach to the neural network by considering it as a non-autonomous dynamical system and analysing it’s velocity vector field which we believe may hold the key to its transparency. Here we study an 81-dimensional version of the Hopfield network, one of the simplest neural models available and using a number of visualisation methods during learning, we elucidate both the mechanisms for the formation of memories in the form of attracting fixed points as well as their 81-dimensional basins of attraction which dictate the decision making process of the network.