ELECTRONIC IMPLEMENTATION OF AN ANALOG ATTRACTOR NEURAL-NETWORK WITH STOCHASTIC LEARNING

We describe and discuss an electronic implementation of an attractor n eural network with plastic synapses. The network undergoes double dyna mics, for the neurons as well as the synapses. Both dynamical processe s are unsupervised. The synaptic dynamics is autonomous, in that it is driven exlusively and perpetually by neural activities. The latter fo llow the network activity via the developing synapses and the influenc e of external stimuli. Such a network self-organizes and is a device w hich converts the gross statistical characteristics of the stimulus in put stream into a set of attractors (reverberations). To maintain for long time the acquired memory, the analog synaptic efficacies are disc retized by a stochastic refresh mechanism. The discretized synaptic me mory has indefinitely long lifetime in the absence of activity in the network. It is modified only by the arrival of new stimuli. The stocha stic refresh mechanism produces transitions at low probability which e nsures that transient stimuli do not create significant modifications and that the system has large palimpsestic memory. A change in the att ractor structure represents a major, macroscopic change in the statist ics of the input stream, which may deform attractors, may create new o nes and may eliminate others. The electronic implementation is complet ely analogue, stochastic and asynchronous. The circuitry of the first prototype is discussed in detail as well as the tests performed on it. In carrying out the implementation we have been guided by biological considerations and by electronic constraints. Both are discussed and n ew insights and lessons for the learning process are proposed.