Predicting model parameters of a spiking neuron network through its local field potential by using deep learning
Year: 2018
Type: dissertation
Abstract: Electrophysiological measurements provide a relatively easy method of probing the\nbrain at the scale of networks. One example of such measurements is the extracellu-\nlar potential that can be obtained by inserting microelectrodes directly into neural\ntissue. The low frequency part of this signal is called the local eld potential (LFP),\nand despite having a solid understanding of the underlying physical principles deter-\nmining these electric potentials, the enormous complexity of neural networks makes\nthe interpretation of the signal di cult. Little is known about exactly what infor-\nmation can be extracted from it.\nConvolutional neural networks, a form of machine learning using deep feed-\nforward arti cial neural networks, have been successfully applied to a wide array\nof tasks, most notably image recognition tasks and natural language processing. In\nthis project, we apply convolutional neural networks to simulated local eld poten-\ntials. We use a newly developed hybrid scheme to simulate the LFP generated by a\nspiking point-neuron network model. To explore what information the LFP signals\ncontain, we address the following question: Can a convolutional neural network be\ntrained to make predictions about the parameters of a spiking point-neuron network\nby using only the LFP it generates?\nWe use a relatively simple model consisting of one excitatory population and\none inhibitory population, and systematically vary three parameters: the excitatory\nsynaptic strength, the relative inhibitory synaptic strength and the amount of ex-\nternal input. We then simulate the LFPs generated by this model network, and use\nthem to train convolutional neural networks to make predictions about the values of\neach parameter. We nd that for this network, it is indeed possible to make fairly\naccurate predictions about the network parameters by using only the LFP it gen-\nerates. These results demonstrate that our approach shows promise for extracting\nnon-trivial information from brain signals, and therefore has the potential to play\nan important role in taking full advantage of these signals in the future.
Source: NVA (Nasjonalt vitenarkiv)
Author: Jan-Eirik W. Skaar
Language: English
Subfield: Cognitive Neuroscience
Field: Neuroscience
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W2902633540
