	US 12,230,021 B2
	System and method for feature visualization in a convolutional neural network
Craig Quiter, Mountain View, CA (US); Siddhartho Bhattacharya, Danville, CA (US); Mayank Ketkar, San Mateo, CA (US); Raluca Musaloiu-Elefteri, San Francisco, CA (US); Wanlin Yang, San Francisco, CA (US); and Sandeep Gangundi, San Jose, CA (US)
Assigned to GM Cruise Holdings LLC, San Francisco, CA (US)
Filed by GM Cruise Holdings LLC, San Francisco, CA (US)
Filed on Apr. 6, 2022, as Appl. No. 17/714,865.
Prior Publication US 2023/0326194 A1, Oct. 12, 2023
Int. Cl. G06V 10/82 (2022.01); G06N 3/08 (2023.01); G06V 10/774 (2022.01)

CPC G06V 10/82 (2022.01) [G06N 3/08 (2013.01); G06V 10/7747 (2022.01)]

20 Claims

1. A computer-implemented method for training a convolutional neural network (CNN), the method comprising:

using a first set of labeled images to train the CNN, the first set of labeled images representing a first object in a plurality of scenes surrounding a vehicle in an environment;

training the CNN to predict a plurality of model errors;

identifying a first model error from the plurality of model errors, wherein the first model error is associated with a type of scene from among the plurality of scenes;

generating a first image by stimulating the CNN, the first image providing a first visualization associated with the first model error;

selecting a second set of labeled images based at least on the first visualization; and

using the second set of labeled images for additional training of the CNN.