Noise, variability and memory formation in spiking neural networks
In order to understand the mechanism driving spiking neural networks (SNNs) to recognize signals on time series; we evolve very small spiking neural networks for recognizing signals in a particular order in presence of noise on membrane potential and variation on silent interval between signals. For recognition of 3 signals in a random input stream the SNNs consist of 3 interneurons and a single output neuron of type adaptive exponential integrate and fire. In addition, the network has 3 dedicated input channels, one for each signal. We use genetic algorithm such that the fitness function rewards for spiking after occurrences of 3 signals in intended order and penalizes spikes elsewhere. We have devised a way to map evolved SNNs on finite state transducer (FST) -- a general model of computation on time series. Furthermore, we demonstrate that SNNs evolved in presence of noise and variation are not only robust to perturbation of neuronal parameters but also emerges a form of memory (thanks to self-excitatory loops -- autapses) such that the network remembers the previous state indefinitely. Finally, we show that evolution may overproduce synaptic connections which can be pruned without impairing performance of the network. The excessive connections are neither important for recognition nor has any role in state maintenance of the network.
|Fri, 26.04.2019||16:30 - 17:00||Panorama|