Computing and Learning with Dynamic Synapses
Wolfgang Maass, Anthony M. Zador · The MIT Press eBooks · 1998
Introduction 1 1 Introduction The models in all other chapters in this book assume that synapses are static, i.e., that they change their "weight" only on the slow time scale of learning. We will discuss in this chapter experimental data which show that this assumption is not justified for biological neural systems. As a matter of fact, this assumption is also unjustified for all hardware implementations of artificial neural nets where the sizes of synaptic "weights" are stored by analog techniques (see Chapter 3). The consequences of this are threefold: i) It is not clear whether implementations of pulsed neural nets in wetware or silicon are able to carry out computations in a way that is predicted by currently existing theoretical models for pulsed neural nets with static synapses. ii) The inherent temporal dynamics of synaptic weights may not just be a curse, but also a blessing: dynamic synapses provide novel computational units for neural comput