US 12,271,816 B2
Class-disparate loss function to address missing annotations in training data
Jasmine Patil, South San Francisco, CA (US)
Assigned to GENENTECH, INC., South San Francisco, CA (US)
Filed by Genentech, Inc., South San Francisco, CA (US)
Filed on Aug. 10, 2022, as Appl. No. 17/885,221.
Application 17/885,221 is a continuation of application No. PCT/US2021/020901, filed on Mar. 4, 2021.
Claims priority of provisional application 62/986,176, filed on Mar. 6, 2020.
Prior Publication US 2022/0383621 A1, Dec. 1, 2022
Int. Cl. G06N 3/08 (2023.01); G06V 10/764 (2022.01); G06V 10/82 (2022.01)
CPC G06N 3/08 (2013.01) [G06V 10/764 (2022.01); G06V 10/82 (2022.01)] 20 Claims
OG exemplary drawing
 
1. A computer-implemented method comprising:
providing a data set including:
an input data element, and
one or more label data portion definitions that each identify a feature of interest within the input data element; and
training a machine-learning model using the data set by performing a set of operations including:
generating one or more model-identified portion definitions that each identify a predicted feature of interest within the input data element, the one or more model-identified portion definitions being generated based on the machine-learning model;
classifying the feature of interest identified by a particular label data portion definition of the one or more label data portion definitions as a false negative by determining a mismatch between the particular label data portion definition and each of the one or more model-identified portion definitions;
classifying the predicted feature of interest identified by a particular model-identified portion definition of the one or more model-identified portion definitions as a false positive by determining a mismatch between the particular model-identified portion definition and each of the one or more label data portion definitions;
calculating a loss using a class-disparate loss function configured to penalize false negatives more than at least some false positives, wherein the calculation includes:
identifying a set of false-positive predicted features of interest each including a predicted feature of interest classified as a false positive;
generating, for each of the set of false-positive predicted features of interest, a confidence metric representing a confidence of the predicted feature of interest existing;
defining a subset of the set of false-positive predicted features of interest based on a quantity of false-positive classifications to be dropped and the confidence metrics; and
assigning a penalty to each of false-positive predicted feature in the subset, wherein the loss is calculated based on the penalties; and
determining a set of parameter values of the machine-learning model based on the loss.