A0620
Title: Improving neural networks interpretability and trustworthiness using polynomials and feature interactions
Authors: Pablo Morala - Universidad Carlos III de Madrid (Spain) [presenting]
Rosa Lillo - Universidad Carlos III de Madrid (Spain)
Inaki Ucar - uc3m-Santander Big Data Institute (Spain)
Jenny Alexandra Cifuentes - Universidad Pontificia Comillas (Spain)
Abstract: Assessing the interpretability and explainability of neural networks is a key factor in adopting their use in a wide range of applications, where their black-box nature poses a problem regarding their trustworthiness. The latest advances in NN2Poly will be presented. NN2Poly is a method that obtains a relationship between a given trained neural network and a polynomial using Taylor expansion and combinatorial properties. This method transforms the nonlinearities in each layer and neuron into polynomials, which can be combined into a final polynomial representation of the neural network. This allows using the polynomial coefficients to interpret the importance of features in the global model, including the interactions between variables up to a certain order. This key aspect is not usually present in other explainability tools. Furthermore, this framework can be used to explore the internal learning process of the neural network by analyzing the obtained polynomials at each layer. Simulations will be presented with both synthetic data (where the interactions are known) and the application to real datasets in regression and classification tasks.
