Homeostatic Plasticity in Recurrent Neural Networks

Recent work in the neurosciences has highlighted the existence of homeostatic plasticity mechanisms that regulate the activity in neuronal networks to a constant level. Functionally similar mechanisms are applied to continuous-time recurrent neural networks. Homeostatic plasticity is found to increase the sensitivity of neural networks to changes in input, giving richer dynamics and robustness to severe changes in the mean level of input. Simulated robots controlled by continuous-time recurrent neural networks using homeostatic plasticity display more complex behaviours than those controlled by non-plastic networks. Synaptic scaling in randomly parameterised networks is observed to generate phototactic behaviour without further adaptation.