David D. Cox

Cox earned his undergraduate degree at Harvard and his doctorate at MIT. His Harvard laboratory studied biological vision while developing computer-vision methods. His CV records Harvard appointments from 2007 to 2019, co-founding DeepHealth in 2017 and becoming IBM director of the MIT-IBM lab in February 2018. The lab’s current name is the MIT-IBM Computing Research Lab.[3][1]

Make model configuration an explicit search problem

Cox coauthored research behind Hyperopt, which treats choices such as model structure and training settings as a search space. A user supplies a performance objective, and an optimization algorithm selects configurations to evaluate. This makes the tuning procedure more systematic and inspectable than relying only on manual trial and error. The chosen objective and evaluation data still determine what the search rewards.[4]

Learn a visual representation by predicting what comes next

In PredNet, Cox and coauthors studied learning from video without labeling every frame. The network predicts future visual input and passes prediction errors through a hierarchy. The authors evaluated what those learned representations captured and how they transferred to other tasks. This is a specific research model for learning from temporal structure, rather than evidence that video prediction alone reproduces human understanding.[5]