Dynamics of neural networks relevant to properties of proteins

Abstract

We studied how the dynamics of Hopfield neural networks depend on computational and physical properties of the network. The dynamics of the network was characterized by the distribution of first passage times (FPT) between the states. The FPT distributions depended on the updating scheme, temperature, connectivity range, and number of stored memories. The FTP distributions were different for synchronous and asynchronous updating, and were more physically consistent for the synchronous than for the asynchronous updating scheme. Neural networks and proteins share common features such as many degrees of freedom, conflicting constraints on energy minimization, and energy functions with many local minima. Thus the general lessons learned here on how the dynamics of neural networks depends on their physical properties may be relevant in understanding how the dynamics of proteins is influenced by similar physical properties.