Robust classification of correlated patterns with a neuromorphic VLSI network of spiking neurons

We demonstrate robust classification of correlated patterns of mean firing rates, using a VLSI network of spiking neurons and spike-driven plastic synapses. The synapses have bistable weights over long time-scales and the transitions from one stable state to the other are driven by the pre and postsynaptic spiking activity. Learning is supervised by a teacher signal which provides an extra current to the output neurons during the training phase. This current steers the activity of the neurons toward the desired value, and the synaptic weights are modified only if the current generated by the plastic synapses does not match the one provided by the teacher signal. If the neuron's response matches the desired output, the synaptic updates are blocked. Such a feature allows the neurons to classify spatial patterns of mean firing rates, even when they have significant correlations. If synaptic updates are stochastic, as in the case of random Poisson input spike trains, the classification performance can be further improved by combining the outcome of multiple neurons together.