US 12,190,565 B2
System and method for domain-agnostic bias reduction with selected sampling for few-shot learning
Ran Tao, Pittsburgh, PA (US); and Marios Savvides, Pittsburgh, PA (US)
Assigned to Carnegie Mellon University, Pittsburgh, PA (US)
Appl. No. 18/267,540
Filed by CARNEGIE MELLON UNIVERSITY, Pittsburgh, PA (US)
PCT Filed Feb. 3, 2022, PCT No. PCT/US2022/015093
§ 371(c)(1), (2) Date Jun. 15, 2023,
PCT Pub. No. WO2022/173650, PCT Pub. Date Aug. 18, 2022.
Claims priority of provisional application 63/148,392, filed on Feb. 11, 2021.
Prior Publication US 2024/0071050 A1, Feb. 29, 2024
Int. Cl. G06K 9/00 (2022.01); G06V 10/764 (2022.01); G06V 10/77 (2022.01); G06V 10/771 (2022.01)
CPC G06V 10/7715 (2022.01) [G06V 10/764 (2022.01); G06V 10/771 (2022.01)] 14 Claims
OG exemplary drawing
 
1. A method of fine-tuning a few shot feature extractor in a classifier trained on a dataset of base classes to reduce biases in novel class feature distribution of the feature extractor caused by an introduction of one or more novel classes, comprising:
inputting few images in each novel class to the feature extractor;
reducing class-agnostic biases in the novel class feature distributions caused by domain differences between the base classes and the novel classes; and
reducing class-specific biases in the novel class feature distribution caused by using only a few samples in the novel classes by selected sampling wherein only samples beneficial to determining the novel class are selected;
wherein class-agnostic bias is feature distribution shifting caused by domain differences between the novel and base classes; and
wherein class-specific bias is the biased estimation resulting from using only a few samples in one class.