Controlling information flow and energy use via adaptive synaptogenesis

Abstract

The adaptive synaptogenesis algorithm is a mathematically defined, random process that, in its present form, creates a feedforward network of excitatory synapses without supervision. The algorithm is fully local and consists of three separate modification processes: random synapse formation, modification of an existing synapse's strength (both strengthening and weakening), and shedding of very weak synapses. The algorithm is shown to have desirable stability properties; further, the algorithm can be parameterized to control the synaptic energy use by a neuron and to control the net information received by a neuron. In addition to the fundamental mathematics on which the algorithm is based, the interaction of parameter settings with characterized random inputs are described. Finally, specific extensions of the algorithm are suggested.