US 11,699,070 B2
Method and apparatus for providing rotational invariant neural networks
Mostafa El-Khamy, San Diego, CA (US); Jungwon Lee, San Diego, CA (US); Yoo Jin Choi, San Diego, CA (US); and Haoyu Ren, San Diego, CA (US)
Assigned to Samsung Electronics Co., Ltd
Filed by Samsung Electronics Co., Ltd., Gyeonggi-do (KR)
Filed on Jun. 25, 2019, as Appl. No. 16/452,005.
Claims priority of provisional application 62/814,099, filed on Mar. 5, 2019.
Prior Publication US 2020/0285894 A1, Sep. 10, 2020
Int. Cl. G06N 3/08 (2023.01); G06N 20/10 (2019.01); G06F 18/214 (2023.01); G06V 10/82 (2022.01); G06V 10/44 (2022.01); G06V 10/46 (2022.01)
CPC G06N 3/08 (2013.01) [G06F 18/214 (2023.01); G06N 20/10 (2019.01); G06V 10/454 (2022.01); G06V 10/462 (2022.01); G06V 10/82 (2022.01)] 18 Claims
OG exemplary drawing
 
1. A method, comprising:
receiving a first input of an image in a first orientation; and
training a kernel to be symmetric, thereby causing an output corresponding to the first input to be identical to an output corresponding to a second input of the image in a second orientation, and
training a convolutional neural network to be symmetric by training multiple kernels at different layers to be symmetric,
wherein the second orientation is a flipped version of the image with respect to the first orientation of the image.