Networks of spiking neurons and plastic synapses: implementation and control

The brain is an incredible system with a computational power that goes further beyond those of our standard computer. It consists of a network of 1011 neurons connected by about 1014 synapses: a massive parallel architecture that suggests that brain performs computation according to completely new strategies which we are far from understanding. To study the nervous system a reasonable starting point is to model its basic units, neurons and synapses, extract the key features, and try to put them together in simple controllable networks. The research group I have been working in focuses its attention on the network dynamics and chooses to model neurons and synapses at a functional level: in this work I consider network of integrate-and-fire neurons connected through synapses that are plastic and bistable. A synapses is said to be plastic when, according to some kind of internal dynamics, it is able to change the “strength”, the efficacy, of the connection between the pre- and post-synaptic neuron. The adjective bistable refers to the number of stable states of efficacy that a synapse can have; we consider synapses with two stable states: potentiated (high efficacy) or depressed (low efficacy). The considered synaptic model is also endowed with a new stop-learning mechanism particularly relevant when dealing with highly correlated patterns. The ability of this kind of systems of reproducing in simulation behaviors observed in biological networks, give sense to an attempt of implementing in hardware the studied network. This thesis situates at this point: the goal of this work is to design, control and test hybrid analog-digital, biologically inspired, hardware systems that behave in agreement with the theoretical and simulations predictions. This class of devices typically goes under the name of neuromorphic VLSI (Very-Large-Scale Integration). Neuromorphic engineering was born from the idea of designing bio-mimetic devices and represents a useful research strategy that contributes to inspire new models, stimulates the theoretical research and that proposes an effective way of implementing stand-alone power-efficient devices. In this work I present two chips, a prototype and a larger device, that are a step towards endowing VLSI, neuromorphic systems with autonomous learning capabilities adequate for not too simple statistics of the stimuli to be learnt. The main novel features of these chips are the implemented type of synaptic plasticity and the configurability of the synaptic connectivity. The reported experimental results demonstrate that the circuits behave in agreement with theoretical predictions and the advantages of the stop-learning synaptic plasticity when highly correlated patterns have to be learnt. The high degree of flexibility of these chips in the definition of the synaptic connectivity is relevant in the perspective of using such devices as building blocks of parallel, distributed multi-chip architectures that will allow to scale up the network dimensions to systems with interesting computational abilities capable to interact with real-world stimuli.