P. Morala Miguélez, J. A. Cifuentes Quintero, R. E. Lillo Rodríguez, I. Ucar Marques
In this work, the method NN2Poly is presented as a tool that can help explain deep artificial feed forward neural networks by means of an alternative representation with polynomials. This extends a previous version of this method that was limited to single hidden layer neural networks. This representation is obtained by using Taylor expansion on the activation functions and combinatorial properties, which allow to obtain the coefficients of a polynomial that performs similarly to a given neural network using its weights. This reduces the number of parameters in the model, improving its interpretability, while also allowing to find interesting connections between neural networks and polynomials. The practical implementation of the method presents some computational limitations that are discussed while also presenting some practical examples.
Keywords: Neural Networks, Interpretability
Scheduled
GT04 Multivariate Analysis and Classification IV. Latest Advances in Explainable Machine Learning
June 7, 2022 6:40 PM
Cloister room