Encoding real values into polychronous spiking networks

Abstract

Spiking neural networks show promising capability in handling the same kind of scaling up of problems as living brains, due to their more faithful similarity to biological neural networks. The big challenge of dealing with spiking neural networks is getting data into and out of them, which requires proper encoding and decoding methods. Presented in this paper is an adaptation of Izhikevich's model of a polychronous spiking network and an encoding scheme for real valued data. Data is chosen arbitrarily to cover the range of the encoding scheme in order to best demonstrate the network's response to different inputs. Preliminary results show that the network is able to recognize distinct input values and respond to them with unique spiking patterns.978-1-4244-8126-2/10/$26.00 ©2010 IEEE

A reversibility analysis of encoding methods for spiking neural networks

“…One way to determine how well they do this is to see if they produce unique bit patterns [12]. Another, and perhaps more telling, method is what has been done here: use the spike patterns to re-generate the original input.…”

“…Further, [11] and [12] demonstrate their respective methods' ability to affect the spiking patterns of a neural network in meaningful ways. The PREM in [10] absolutely gets information through its input neurons, as the MLE decoding can recover it only once the neurons' spiking patterns have been read.…”

“…The range and resolution of this encoding method is strictly determined by the number of bits used. In this paper as in [12], 12 bits (and thus 24 neurons) are used.…”

“…10 shows the performance of the dual-neuron n-bit representation encoding method on the multi-frequency sine function given by (5). The resolution provided by the 12 bits used in [12] allows for 4096 discrete values to be encoded, which allows for a range of slightly more than [-20, 20] to be covered by the encoding mechanism with a resolution of two decimal places. The amplitude of the input sine wave is thus set to 20 to fully explore the range of this encoding method.…”

“…The three encoding methods studied in this paper are a Poisson rate encoding method (PREM) [10], Gaussian receptive fields (GRF) [11], and a dual-neuron n-bit representation (DNNR) [12]. PREM and GRF are each kinds of temporal encoding.…”