Robustness of learning of deep and shallow networks
Specialist supervisor: RNDr. Petra Vidnerová, Ph.D.
Deep networks became state-of-the-art performance on a variety of pattern-recognition tasks, most notably visual classiffication ones. Several recent studies revealed that some deep networks can be easily fooled by changing an image in a way imperceptible to humans which can cause a deep network to label the image as something else entirely. Thus it is important to investigate robustness of learning of networks with respect to adversarial patterns.
The goal of the dissertation is to design multiobjective algorithms generating imperceptible perturbations of images leading to their misclassifications. Evolutionary algorithms for generation of adversarial patterns will be developed that maximize the error of the model while being as similar to original training patterns as possible. The role of depth of networks and locality properties of computational units in misclassiffication will be analyzed. Methods of making neural networks more robust to such adversarial patterns will be investigated.