Abe Davis

Most recent advances in AI and visual computing have been driven by techniques that favor learning from large datasets over more direct applications of domain knowledge. Such techniques can be incredibly powerful, but also less interpretable and harder to predict, complicating many of the abstractions that have supported progress in computer science for the past several decades. In this talk, I will discuss how this has created significant new opportunities for research that blends technical innovation design methodologies to tackle problems at the intersection of different disciplines.

New Opportunities for Interaction Design: I will discuss some of my group’s work that integrates learning and optimization into the design of interactive tools. More specifically, looking at interactive tools that take new approaches to human-in-the-loop learning.

Custom, Precise, & Efficient Data Capture: Abundant sources of data tend to have very biased sampling, which limits the impact of data-driven methods in domains that depend on scarce or highly specific distributions of data—e.g., applications in healthcare, infrastructure, and the sciences. I’ll discuss my group’s work on using mobile devices to more easily deploy and scale customizable data collection protocol in uncontrolled environments.

Video Forensics and Noise-Coded Illumination: Video forensics in completely general settings (i.e., the problem of determining if a completely arbitrary video is fake or manipulated) is almost hopeless in the long term. However, if we approach forensic analysis as a design problem and allow ourselves to build systems that address more specific settings, we can develop solutions with real long-term practical potential. I will discuss my group’s work on noise-coded illumination (NCI) as one example of this where, by encoding subtle watermarks into the illumination of a given environment, we can create a significant information advantage for forensic verification of video captured in that environment.