Machine and Deep Learning Seminar

Abstract – A Novel Perspective on Robustness in Deep Learning 

Despite their unprecedented success, neural networks exhibit different vulnerabilities when facing an adversary which should be addressed before deploying them in safety-critical applications. In this thesis, we study these issues to understand their nature better, aiming to make neural networks more robust. We exploit various statistical tools to describe these adversaries and shed light on their behavior. Using these insights, we revisit robustness from novel perspectives and propose efficient training methods that work well against a suite of adversarial threats.