Computation with Spiking Neurons

We develop two artificial neural network models which use ‘spiking’ neurons to perform recognition and memory tasks. The first task studied is that of recognising spoken words. We present a network built from approximately 1000 biologically plausible integrate-and-fire neurons which was developed as the winning solution to an academic competition posed by John Hopfield and Carlos Brody. The network recognises a vocabulary of ten words, regardless of the speed at which they are spoken (‘time-warp invariance’). The network employs two key mechanisms: transient equality in the firing rates of subsets of neurons, and the inducement of synchrony among groups of weakly-connected neurons. In the remainder of the thesis we study an autoassociative memory. The network consists of a single layer of spiking neurons with recurrent time-delay connections. Each neuron is built from a number of coincidence detector subunits, and can be thought of as an approximation to a binary ‘sigma-pi’ unit. The network stores and recalls spatiotemporal spike sequences (patterns of neural activity over time). When stimulated with a noisy or incomplete version of a stored memory, the network will ‘clean up’ the memory and, if periodic spike sequences are used, repeatedly replay the corrected, completed memory. Unlike most memory models, the system is able to simultaneously recall more than one stored memory, using a single population of neurons. We explore the capacity properties of this ‘Concurrent Recall Network’. Using a mapping from real numbers to spatiotemporal spike sequences, we extend the Concurrent Recall Network into a line attractor memory. In this mode of operation, it is able not only to store individual memories (point attractors) but also a continuum of memories (a line attractor). After training, the network can stably recall the spike sequence corresponding to any point along the attractor. Furthermore, by applying a global ‘velocity control’ signal to all neurons in the network, we can cause the recalled state to drift along the attractor in a controlled fashion. Finally, we demonstrate the ability of the Concurrent Recall Network to store multiple line attractors, recall different positions along each attractor concurrently and independently move along each attractor.