Systematic Approach to Increase Robustness of Feature Vectors for Neural Networks

摘要

Generalisation is one common problem that all supervised machine learning problems must overcome. In the supervised learning context, a deep neural network consisting of multiple vectors of features extracted from the data via many layers arc used for various classification, detection, and regression applications. The ability for a deep neural network to generalise is crucial in many applications and having more robust feature vectors provide the desired results in a consistent fashion given variations in the data during inference is required. Given a deep neural network architecture, there exist many approaches in increasing robustness of the feature vectors .such as data augmentation, but it often takes a lot of resources and time to do a full-factorial search to find the optimal tuning for data augmentation without knowing exactly what tuning helps how much due to the complex architecture of deep neural networks. Furthermore, finding sub-optimal parameters due to not running full-factorial search can end up in costly manual labors to correct mis-classifications in the case of auto-data annotation or lead to missed detections during safety-critical operations such as autonomous vehicles. By utilizing statistical analysis via orthogonal array, the data augmentation for deep neural networks can be effectively analyzed, and the optimal set of parameters can he found in a short amount of time.