Supervised learning in spiking neural networks with ReSuMe method

Filip Ponulak

Full text: http://d1.cie.put.poznan.pl/dav/fp/Ponulak_PhD_Thesis.pdf

Find on Google Scholar

---

Abstract

Supervised learning in Spiking Neural Networks (SNN) is considered in this dissertation. Spiking networks represent a special class of artificial neural networks, in which neuron models communicate by sending spikes (action potentials).
SNN are investigated here in the context of their potential applications to movement control in neuroprostheses, i.e. in the systems that aim to substitute or restore locomotory functions in human subjects with the movement disorders.
From the control point of view, human limb is a multidimensional, nonlinear, nonstationary object, characterized by the rich embedded dynamics. This is the reason for the low effectiveness of the traditional control approaches in neuroprosthesis. One of the new concepts to solve this problem is to mimic functions of the Central Nervous System by the SNN-based controllers. However, the analysis of the recent supervised learning methods for SNN revealed that the existing algorithms are not suitable for the task at hand. Investigation on the new approaches in this issue led the author of this dissertation to the invention of a novel supervised learning method, called ReSuMe (Remote Supervised Method).
Results of the extensive simulation studies presented in this dissertation confirm that the networks of spiking neurons can be efficiently trained with ReSuMe to store and precisely reproduce the spatiotemporal patterns of spikes. Convergence of the ReSuMe learning process was formally proved for the spiking neurons excited with a single pair of input-teacher spikes. Thorough parametric analysis of the learning properties and the memory capabilities of SNN was performed and the generalization property of the spiking neurons trained with ReSuMe was demonstrated. ReSuMe was successfully applied to control movement of a simple leg model. A new, SNN-based model of the programmable
Central Pattern Generator (CPG) was proposed and tested.
Conclusions presented in this dissertation are formulated based on the results of about 6000 computer simulations. Beside the simulation studies, the ReSuMe method was implemented in the FPGA system. The hardware platform is intended to the future applications of spiking networks in the real-life tasks.

---

---

Ratings & reviews

• Supervised learning in spiking neural networks with ReSuMe method

  Anonymous reviewer

  Anonymous

  100 88 0