A learning rule balancing energy consumption and information maximization in a feed-forward neuronal network

Abstract: Information measures are often used to assess the efficacy of neural networks, and learning rules can be derived through optimization procedures on such measures. In biological neural networks, computation is restricted by the amount of available resources. Considering energy restrictions, it is thus reasonable to balance information processing efficacy with energy consumption. Here, we studied networks of non-linear Hawkes neurons and assessed the information flow through these networks using mutual information. We then applied gradient descent for a combination of mutual information and energetic costs to obtain a learning rule. Through this procedure, we obtained a rule containing a sliding threshold, similar to the Bienenstock-Cooper-Munro rule. The rule contains terms local in time and in space plus one global variable common to the whole network. The rule thus belongs to so-called three-factor rules and the global variable could be related to a number of biological processes. In neural networks using this learning rule, frequent inputs get mapped onto low energy orbits of the network while rare inputs aren't learned.