US 12,554,992 B2
Self-supervised learning using in-painting
Naomi Simumba, Yokohama (JP); Michiaki Tatsubori, Oiso (JP); and Daiki Kimura, Midori-ku (JP)
Assigned to International Business Machines Corporation, Armonk, NY (US)
Filed by INTERNATIONAL BUSINESS MACHINES CORPORATION, Armonk, NY (US)
Filed on May 22, 2023, as Appl. No. 18/321,143.
Prior Publication US 2024/0394547 A1, Nov. 28, 2024
Int. Cl. G06N 3/0455 (2023.01); G06N 3/0895 (2023.01); G06T 5/70 (2024.01); G06T 5/77 (2024.01)
CPC G06N 3/0895 (2023.01) [G06N 3/0455 (2023.01); G06T 5/70 (2024.01); G06T 5/77 (2024.01); G06T 2207/10032 (2013.01); G06T 2207/20081 (2013.01)] 20 Claims
OG exemplary drawing
 
1. A method using a computing device for self-supervised learning of a machine learning model using images, the method comprising:
masking by the computing device a section of each of multiple images made available from an image database containing one or more images to generate a respective partially masked image of each of the multiple images;
encoding by an autoencoder associated with the computing device each partially masked image to generate one or more encodings respectively representing each partially masked image;
decoding by the autoencoder each of the one or more encodings into one or more decoded encodings as respective first reconstructions of the multiple images;
comparing by the computing device each of the one or more decoded encodings of each partially masked image with a corresponding image of the multiple images to generate a reconstruction loss;
augmenting by the computing device each partially masked image according to a data augmentation policy to generate one or more augmented partially masked images;
inputting by the computing device each augmented partially masked image into the autoencoder to obtain an augmented model output;
determining by the computing device a total loss by using:
the reconstruction loss and
a consistency loss of the augmented model output; and
improving by the computing device the autoencoder based upon the determined total loss to create an improved autoencoder.