Deconstructing a Neural Network

ISBN: 978-981-08-0693-4_CGAT2

Authors: Patrick Moghames and Bikramjit Banerjee

Abstract:

We focus on the problem of constructing an equivalent function for a given neural network, when the training/test data are not available. The unknown function captured by the neural net is represented as the Taylor series expansion in terms of the inputs, and the relevant coefficients are computed from the weights of the network. We argue that such deconstruction of a neural network can be a useful tool in complexity reduction.

Keywords: Neural Network, Taylor Series, Approximation