Adversarially trained smooth classifiers reach provably robust accuracy

Artificial intelligence has evolved to become a revolutionary technology. It is rapidly changing the economy, both by creating new opportunities (it's the backbone of the gig economy) and by bringing venerable institutions, like transportation, into the 21st century. Yet deep at its core something is amiss, and more and more experts are worried: the technology seems to be extremely brittle, a phenomenon epitomized by adversarial examples. Adversarial examples exploit weaknesses in modern AI. Today, most successful AI applications use machine learning (more specifically, supervised learning) by training big neural networks to mimic input-output mappings on sample data.