We test whether the accepted ways of evaluating AI capture how these systems behave once they leave the benchmark, with current work in medical imaging and neuroscience.
Each project starts from one question: does the standard metric measure what we care about? Select a bubble to explore an area.
An alternative to the AP poll for the Power Four and Notre Dame. Opponent-adjusted ratings, a strength of schedule that re-grades itself every week, and what-if schedules you can play out on a pin board of red string.
Using the National Lung Screening Trial, we compared AI segmentation on the same CT acquisition reconstructed with two different filters. Overlap scores did not change. Measured tumor volume did.
Box-prompt result is exploratory. Protocol fixed before training; test set scored once.
A spiking model of the fruit fly connectome, given a simple body and placed in a hide and seek game. Silencing two dozen neurons in its compass pathway leaves the seeker almost unable to win.
Publication pendingBrain wave classification and Hurst exponent analysis of long-range structure in EEG recordings.
Publication pendingA detector for AI-generated video that checks visual, temporal, and audio evidence separately, explains each flag, and attributes the clip to its likely generator.
Publication pendingRead about the lab and how to collaborate.
About the labUpdates on publications and new projects.
linkedin.com/in/jamirwin