Intrinsic plasticity in autonomous recurrent neural networks

Abstract

Event Abstract Back to Event Intrinsic plasticity in autonomous recurrent neural networks Dimitrije Markovic1* and Claudius Gros1 1 J. W. Goethe University, Institute for Theoretical Physics, Germany We investigate generic properties of neural networks governed by autonomous self-regulation using a previously proposed model of intrinsic plasticity [1, 2]. The learning rules for the adaptation of _ring-rate neurons were derived from information-theoretical prin-ciples. We show that this polyhomeostatic optimization, which in contrast to homeostatic regulation aims at stabilization of a speci_c target distribution of neural activities, gives rise to non-trivial dynamical states when recurrent interactions are introduced. We consider individual self-coupled neurons and networks of _ring-rate neurons, with intrinsic plasticity trying to achieve maximal en-tropy of output _ring rates. We _nd, in particular, that the intro-duction of intrinsic adaption leads to a destruction of all attractors in Hop_eld-like network setup [3]. Surprisingly, for large networks, we observe either chaotic or intermittent bursting behavior [4], de- pending on the speci_ed average _ring rates. A property similar to networks of spiking neurons. Interestingly, both states are globally attracting in their respective phase spaces. Autonomous and poly-homeostatic self-control of dynamical systems may lead in general to non-trivial dynamical states and novel phenomena.