Enhancing Neural Network Robustness Through Intelligent Data Transformation

This paper exploits the origins of artificial neural networks (ANN s) based on approximation theory, to refocus efforts on data transformation, and away from the exclusive attention on the development of more and more complex architectures. Machine learning (ML) networks efficiently accept raw input data and determine their own features for the training process. Features are chosen to optimize accuracy, but typically are not designed to be robust, leaving the network susceptible to noise. Thus, a case is made for the reintroduction of data transformation techniques, through an intelligent algorithm, to enhance the robustness of the ML networks. In a sense, the science and mathematics of AI algorithms must provide a path for the accurate, reliable, and rapid deployment of intelligent algorithms without an over-reliance of technological prowess such as inexpensive bandwidth and computing power. We developed an algorithm for image classification based on intelligent geometric transformations, which forms the input data for any neural network architecture of choice. Our results show that adopting intelligent data transformations for pre-processing network input leads, in comparison to a conventional raw image input convolutional neural network, to a more robust model to both random and adversarial noise, offering network designers enhanced control over the trade-off between accuracy and robustness.