Dr. Boult will talk about some of the issues with machine learning (which some call AI) in an open world.  In a closed world, all possible classes/choices for the system are known ahead of time. The closed world assumption is made by the vast majority of ML/AI systems.  When an unexpected input is encountered the system will map it to one of its known choices.  Even if the system is trained to reject millions of “uninteresting” inputs, the system still gets the majority of inputs not seen in training wrong.   We’ll briefly discuss how we have been addressing this over the past decade.

A special type of “unknown” input is what is called adversarial perturbations.  In these inputs, the system is provided input with nearly imperceptible changes in the values.  These small changes are designed so humans see one thing, yet computers using deep networks see something very different.  In my image used to advertise this talk, we generated special perturbations and added them to the image such that the leading commercial face recognition systems in 2018 (AWS, Microsoft), as well as multiple government systems, all say the image is Barack Obama!!!  Despite almost a decade of research on these adversarial perturbations, no accepted theory explains why they exist and no defense is good at protecting against them if the attacker gets access to the system.   This is true even if the system owners train their ML system with millions of perturbations.

Many people like to hype current “Artificial Intelligence” and speculate about things like AI consciousness.  I’ll end the talk by discussing this and that in my view we are not even up to Artificial Stupidity though many systems work well in very narrow domains — maybe we are approaching an artificial autistic savant.