US 11,727,275 B2
Model training method and apparatus
Hwidong Na, Seongnam-si (KR); Hyohyeong Kang, Hwaseong-si (KR); Hogyeong Kim, Daejeon (KR); and Hoshik Lee, Seongnam-si (KR)
Assigned to Samsung Electronics Co., Ltd., Suwon-si (KR)
Filed by Samsung Electronics Co., Ltd., Suwon-si (KR)
Filed on Sep. 22, 2022, as Appl. No. 17/950,521.
Application 17/950,521 is a continuation of application No. 16/562,783, filed on Sep. 6, 2019, granted, now 11,488,013.
Claims priority of application No. 10-2019-0053131 (KR), filed on May 7, 2019.
Prior Publication US 2023/0035351 A1, Feb. 2, 2023
This patent is subject to a terminal disclaimer.
Int. Cl. G06N 3/08 (2023.01)
CPC G06N 3/08 (2013.01) 16 Claims
OG exemplary drawing
 
1. A method of training a model, the method comprising:
acquiring a recognition result of a teacher model and a recognition result of a student model for an input sequence;
determining an adversarial loss based on a degree to which an output sequence of the teacher model and an output sequence of the student model that are respectively output as recognition results for the input sequence are distinguished from each other; and
training the student model to reduce the adversarial loss,
wherein the determining of the adversarial loss comprises determining the adversarial loss by applying a Gumbel-max based on a probabilities of elements included in the output sequence of the teacher model.