From continous behaviour to discrete knowledge

Abstract

Neural networks have proven to be very powerful techniques for solving a wide range of tasks. However, the learned concepts are unreadable for humans. Some works try to obtain symbolic models from the networks, once these networks have been trained, allowing to understand the model by means of decision trees or rules that are closer to human understanding. The main problem of this approach is that neural networks output a continuous range of values, so even though a symbolic technique could be used to work with continuous classes, this output would still be hard to understand for humans. In this work, we present a system that is able to model a neural network behaviour by discretizing its outputs with a vector quantization approach, allowing to apply the symbolic method.