Imaging Science Seminar: Classification and calibration in low data regime
Classification and calibration in low data regime
Dr. Yi Yao
Sr. Technical Manager
Center for Vision Technologies
SRI International
Abstract:
Few-shot learning (FSL) remains challenging due to the longstanding difficulties of (1) large semantic gaps between the base and novel classes and (2) sparse and non-uniformly distributed labeled samples from novel classes. We propose innovative solutions to tackle these two challenges and demonstrate state-of-the-art performances on five FSL datasets. Meanwhile, existing calibration algorithms address the problem of covariate shift via unsupervised domain adaptation. These methods require unlabeled data from the target domain, which may not be available at the stage of calibration in real-world applications. Moreover, their performances heavily depend on the disparity between the distribution of the source and target domains. To address these two limitations, we present novel calibration solutions and show decreased expected calibration errors on the Office-home dataset.
Speaker Bio:
Yi Yao (Ph.D.), Sr. Technical Manager at the Center for Vision Technologies, SRI International, has significant experience in computer vision, machine learning, data mining, social media exploitation, visual surveillance, object recognition, event recognition, and object tracking. She has 50+ papers in peer-reviewed journals and conferences, and 10+ granted patents. Dr. Yao is the PI/co-PI for DARPA LwLL: learning with limited training data; NGA Ship Detection: multi-modal small ship detection from satellite imagery; DARPA Providence: in-situ continual learning using streaming data with noisy annotation, class imbalance, and adversarial/out-of-distribution samples; previously IARPA DIVA: complex activity recognition from complex scenes and distributed cameras; IARPA PMSIA: economic indicator prediction using raw measurements from satellite imagery; NGA BIG/BAD AC: complex geospatial entity detection from satellite imagery; and CTTSO GRI-FN: automated mobile geolocation in GPS-denied environment. Technical Advisor for DARPA SAIL-ON, DARPA XAI, and IARPA TrojAI.
Intended Audience:
Beginners, undergraduates, graduates, experts. Those with interest in the topic.
Event Snapshot
When and Where
Who
Open to the Public
Interpreter Requested?
No