Retrospective processing of video data captured at the edge typically requires creation of a DNN that is custom-built through transfer learning for the task at hand. Creating such a DNN requires construction of a large training set of examples, typically 10^3 to 10^4 true positives. In cases when a target is rare, a domain expert needs to go through millions of potential images to find a few positive instances of the target. Transmitting such a tremendous volume of data to the cloud for inspection is infeasible due to limited network bandwidth and significant cost.
My thesis will address these challenges. In particular, I claim that it is feasible and effective to create a human-in-the-loop labeling system that performs time-critical, distributed, low base rate active learning in an edge computing setting for novel search queries. To this end, I will present a prototype implementation of a system Delphi, an interactive labeling system that performs machine learning in the background. I will describe how Delphi can be used by a domain expert with no machine learning knowledge to customize and serve models for a novel target, while conserving network and human attention bandwidth. In this work, I will be validating the usability and generality of Delphi and explore building a tuning-wizard that can select the optimal learning and server parameters to match network bandwidths, data distribution, and computational loads on cloudlets.
Mahadev Satyanarayanan (Chair)
Padmanabhan Pillai (Intel Labs)
Zoom Participation. See announcement.