The SGE framework: Discovering spatio-temporal patterns in biological systems with spiking neural ne
暂无分享,去创建一个
Developing smart machines that are able to recognize patterns is an active area of engineering research and has been for some time. In the 40s McCulloch and Pitts developed artificial neural network models which have been greatly advanced to give us algorithms that can recognize or learn patterns. However, much of this work focused on separating patterns that are stationary in time and space. While this has been valuable in many engineering applications, it fails to capture the complex adaptive nature of living systems that are neither stationary in time or space. The goal of this dissertation is to advance our ability to model complex adaptive systems including both spatial and temporal characteristics and apply them to engineering problems.
While several spatio-temporal pattern recognition approaches exist, most transform the problem into the static domain. Others utilize more biologically plausible models such as spiking neural networks which have the advantage of precisely incorporating spatio-temporal patterns. Current approaches using these models proceed primarily with traditional training methods that update only the network weights or exhaustive optimization techniques. However, to date, a mechanism does not exist to fully utilize these models with respect to learning patterns. Evolutionary computation has been suggested as a means to explore the full parameter space of these models but to date, this has not been accomplished.
This dissertation lays out the framework, along with specific examples, for training spiking neural networks using genetic algorithms and expert knowledge. We developed a spiking neural network simulator which includes leaky integrate and fire neurons, dynamic synapses and Hebbian long-term synaptic plasticity. By coupling this simulator to the CHC genetic algorithm with a tunable fitness function we successfully trained a spiking neural network to recognize patterns from a temporal XOR and a tonic burster. We extended this, in combination with experimental neural recordings, to develop a testable model of taste processing in the nucleus of the solitary tract in the rat.
We believe that this framework provides the first example of an adaptable methodology to solve spatial temporal pattern recognition problems using artificial spiking neural network models.