US 12,394,182 B2
Systems and methods for multi-teacher group-distillation for long-tail classification
Tanvir Mahmud, Austin, TX (US); Chun-Hao Liu, Fremont, CA (US); and Burhaneddin Yaman, San Jose, CA (US)
Assigned to Robert Bosch GmbH, (DE)
Filed by Robert Bosch GmbH, Stuttgart (DE)
Filed on Sep. 21, 2022, as Appl. No. 17/949,504.
Prior Publication US 2024/0096067 A1, Mar. 21, 2024
Int. Cl. G06V 10/774 (2022.01); G06N 3/08 (2023.01)
CPC G06V 10/774 (2022.01) [G06N 3/08 (2013.01)] 20 Claims
OG exemplary drawing
 
1. A computer-implemented method of classifying a long-tail distribution of data, the method comprising:
receiving data deriving from one or more sensors;
classifying the data into a plurality of classes, wherein the step of classifying uses (i) a feature-extractor backbone model configured to extract features from the data and (ii) a classifier model configured to classify the data based on the extracted features;
grouping the plurality of classes of data into a plurality of groups, wherein each group includes a number of the classes, and wherein the number of classes in each group is controllable;
training a plurality of teacher models to predict classes, wherein each teacher model is trained with (i) the data in a respective one of the groups, and (ii) the feature-extractor backbone model; and
after the step of training, merging outputs of the plurality of teacher models into a final class prediction model configured to classify the data.