Adaptive Chaotic Map Selection to Reduce Overfitting in Artificial Neural Networks

Artificial neural networks (ANNs) have become a widely used machine learning tool in many different and broad fields of application. ANNs may be trained from training examples to perform various prediction or classification tasks. However, one of the problems exhibited by ANNs is that they typically overlearn the training examples and perform poorly on new inputs they haven’t been trained on. As a result, the ANN does not generalize well to novel data. The neuroscience field has shown some evidence that human cognition exhibits some form of chaotic behaviour. In our research, we use a mathematical tool called a chaotic strange attractor to generate chaotic sequences of values to inject into an ANN to reduce overfitting. While there are many different kinds of chaotic strange attractors, we propose a novel method of adaptively choosing a single or a subset of strange attractor(s), based on the nature of the training data, as well as on the nature of the distribution of possible solutions in the solution space. To assess the efficacy of our method, we propose to use three common data sets for training and testing and we propose to compare our method to the baseline ANN, dropout (DO)

Faculty Supervisor:

Ken Ferens;Witold Kinsner

Student:

Partner:

Canadian Tire Corporation

Discipline:

Engineering

Sector:

Retail trade

University:

University of Manitoba

Program: