Adaptive chaotic injection to reduce overfitting in artificial neural networks
Abstract
Artificial neural networks (ANNs) have become an integral tool in various fields of
research. ANNs are mathematical models which can be trained to perform various prediction
tasks. The effectiveness of an ANN can be impacted by overfitting which occurs when the ANN
overfits to the training data. As a result, the ANN does not generalize well to novel data. In our
research, we assess the feasibility of using a chaotic strange attractor to generate sequences of
values to inject into an ANN to reduce overfitting. An adaptive method was developed to scale
and inject the values into the neurons throughout training. The chaotic injection (CI) was tested
on three benchmark datasets using different ANN models. The results were compared against the
baseline ANN, dropout (DO), and Gaussian noise injection (GNI). The CI improved the
performance of the ANN and converged faster than DO and GNI.