US 12,079,738 B2
Variance of gradient based active learning framework for training perception algorithms
Armin Parchami, Ann Arbor, MI (US); Ghassan AlRegib, Johns Creek, GA (US); Dogancan Temel, Atlanta, GA (US); Mohit Prabhushankar, Atlanta, GA (US); and Gukyeong Kwon, Atlanta, GA (US)
Assigned to Ford Global Technologies, LLC, Dearborn, MI (US)
Filed by Ford Global Technologies, LLC, Dearborn, MI (US)
Filed on Feb. 10, 2021, as Appl. No. 17/172,854.
Prior Publication US 2022/0253724 A1, Aug. 11, 2022
Int. Cl. G06N 5/04 (2023.01); G06F 18/214 (2023.01); G06F 18/24 (2023.01); G06N 3/02 (2006.01); G06N 3/08 (2023.01); G06N 3/084 (2023.01); G06N 20/00 (2019.01)
CPC G06N 5/04 (2013.01) [G06F 18/214 (2023.01); G06N 20/00 (2019.01); G06F 18/24 (2023.01); G06N 3/02 (2013.01); G06N 3/08 (2013.01); G06N 3/084 (2013.01)] 20 Claims
OG exemplary drawing
 
1. A system for training models using active learning, the system comprising:
a sensor;
one or more processors;
a system memory, the system memory storing instructions to cause the one or more processors to:
receive a dataset, the dataset comprising a plurality of data frames;
classify an object in the dataset based on a machine learning model;
calculate a plurality of loss values, each of the loss values characterizing a discrepancy between the classification of the object in the dataset and one of a plurality of potential classifications of that object;
generate a gradient for each of the plurality of loss values;
add corresponding gradients to a gradient pool;
calculate a variance of the gradient pool; and
request annotation of at least a portion of the dataset based in part on the variance of the gradient pool.