Abstract: We examine the relationship between the performance of a neural network (NN) model in recognizing emotions and the fuzziness of the rules it implements. We first propose a NN implementation of the Mimic system, which was designed to code facial expressions. It consists of rules mapping the movements of facial areas to the activation of facial muscles, and uses those muscle activations to predict the emotion expressed by the face. The suitability of the NN for extracting its underlying rules is assessed by a measure of the fuzziness on the set of synaptic weights. Preliminary simulations indicate that improving performance by training only the first layer of weights is done at the expense of rule clarity; however, sharpening the weights at their fuzziest point leads to perfect rule extractability and almost perfect recognition. By contrast, limiting the learning to the second layer of weights leads to perfect recognition with limited potential for rule extraction.

Key-Words: Hybrid systems, Neural Networks, Rule extraction, Fuzziness, Facial expressions, Emotion recognition.