Abstract
In a recent study, we demonstrated that a simple two-layer network of spiking neurons equipped with a novel Spike-Time Dependent Plasticity rule was capable of spontaneously learning to detect complex dynamic visual stimuli (Bichler et al, 2012, Neural Networks, 32, 330-48). The network received spike-like events from a Dynamic Vision Sensor chip that asynchronously generates "ON" and "OFF" spiking events in response to changes in luminance. After a few minutes of stimulation corresponding to the passage of cars on a 6-lane freeway, we found that the 60 neurons in the first layer had formed receptive fields that correspond to car-like shapes at particular locations on the road, whereas the 10 neurons in the second layer had learned to "count" cars going by on each of the six lanes. Importantly, this sort of learning was entirely unsupervised and reflected the fact that the STDP leads neurons to become selective to spatio-temporal spike patterns that occurred repeatedly. In general, a few tens of repetitions are enough for the selectivity to develop. In the current study, we demonstrated that this sort of learning could effectively lead to the development of "grandmother cell" coding. For example, a neuron that initially responded to continuous random synaptic inputs at roughly 2 spikes/second could be trained to respond selectively to a specific repeating sequence after only a few tens of repetitions. The remarkable finding is that when we returned to the original random input pattern, the neuron was now completely silent and would remain silent indefinitely. However, if ever the pattern used during training is presented again, even after a very long delay, the neuron would immediately respond at a short latency. We propose that this simple mechanism could underlie the ability of the brain to store long-term sensory memories.
Meeting abstract presented at VSS 2013