US 12,135,762 B1
Extended semi-supervised learning generative adversarial network
Andrew W. Hahn, Ellicott City, MD (US); Murali Tummala, Monterey, CA (US); and James Scrofani, Warrenton, VA (US)
Assigned to THE UNITED STATES OF AMERICA, AS REPRESENTED BY THE SECRETARY OF THE NAVY, Arlington, VA (US)
Filed by THE UNITED STATES OF AMERICA, AS REPRESENTED BY THE SECRETARY OF THE NAVY, Arlington, VA (US)
Filed on Dec. 7, 2020, as Appl. No. 17/113,731.
Claims priority of provisional application 63/111,393, filed on Nov. 9, 2020.
Claims priority of provisional application 62/948,460, filed on Dec. 16, 2019.
Int. Cl. G06N 20/00 (2019.01); G06F 18/20 (2023.01); G06F 18/2132 (2023.01); G06F 18/214 (2023.01); G06F 18/243 (2023.01)
CPC G06F 18/2132 (2023.01) [G06F 18/2155 (2023.01); G06F 18/24317 (2023.01); G06N 20/00 (2019.01)] 13 Claims
OG exemplary drawing
 
1. An extended semi-supervised learning (ESSL) generative adversarial network (GAN) comprising:
a generator network G, the generator network G comprising:
an input for receiving a noise vector z;
an input for receiving a conditional vector c, wherein the conditional vector c is the size K, and K is a number of classes, wherein the number of classes is based on priori knowledge of the number of classes of a generator loss function LD; and
an output for outputting a synthetic data batch G(z|c) generated by the generator network G, wherein the synthetic data batch G(z|c) is based on a distribution of a real signal batch x and one or more classes of the real signal batch x; and
a discriminator network D communicatively coupled to the generator network G, the discriminator network D comprising:
an input for receiving the synthetic data batch G(z|c);
an input for receiving the real signal batch x;
discriminator weights;
an optimizer for optimizing the synthetic data batch and the real data batch as its appropriate class; and
an output for outputting an estimated label vector γ generated by the discriminator network D for each synthetic data batch G(z|c) input and each real signal batch x input based on the discriminator weights.