Parametric Properties in Chaotic Neural Networks

Keywords
Loading...
Thumbnail Image
Issue Date
2016-08-25
Language
en
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
In the framework of reservoir computing a randomly constructed recurrent neural network (RNN) is used to function as a reservoir containing input history and its many random transformations. Output units transform whatever is inside this reservoir into a meaningful output. However, it is di cult to appropriately construct a random reservoir such that the complete model is able to learn and perform a certain task. A model's performance depends on two things: the parameters of the RNN and the task's complexity. The used network parameters need to be ne-tuned to the task at hand for achieving the best results. First we propose and investigate several measures of quantifying the di culty of the chosen to-be-trained pattern. The used patters are visual. Hereafter we investigate RNN performance after FORCE learning on tasks of di erent complexities, selected with the apparently most appropriate of these measures, and we investigate how an RNN's parameters determine its eventual capability to learn. Under investigation are reservoir size, connectivity and weight scaling. All ndings were done through simulation studies, with additional theoretical explanations and references to relevant literature.
Description
Citation
Faculty
Faculteit der Sociale Wetenschappen