US 12,242,971 B2
Adversarial training of machine learning models
Xiaodong Liu, Bellevue, WA (US); Jianfeng Gao, Woodinville, WA (US); Pengcheng He, Sammamish, WA (US); and Weizhu Chen, Kirkland, WA (US)
Assigned to Microsoft Technology Licensing, LLC, Redmond, WA (US)
Filed by Microsoft Technology Licensing, LLC, Redmond, WA (US)
Filed on Jan. 29, 2020, as Appl. No. 16/775,635.
Claims priority of provisional application 62/932,324, filed on Nov. 7, 2019.
Prior Publication US 2021/0142181 A1, May 13, 2021
Int. Cl. G06N 3/088 (2023.01); G06N 3/045 (2023.01)
CPC G06N 3/088 (2013.01) [G06N 3/045 (2023.01)] 23 Claims
OG exemplary drawing
 
1. A method performed on a computing device, the method comprising:
obtaining a machine learning model having one or more embedding layers and pretrained parameters;
performing a tuning stage on the machine learning model by using labeled training samples to tune the pretrained parameters, the tuning stage including:
performing noise adjustment by adding noise to embeddings produced by the one or more embedding layers that represent text of the labeled training samples to obtain noise-adjusted embeddings, and
adjusting the pretrained parameters to obtain adapted parameters, the adjusting being based at least on a difference between first output distributions determined by the machine learning model using the embeddings representing the text of the labeled training samples and second output distributions determined by the machine learning model using the noise-adjusted embeddings, wherein the first output distributions and the second output distributions are output by the machine learning model and include different values corresponding to different likelihoods that a particular labeled training sample has a particular label; and
outputting a tuned machine learning model, the tuned machine learning model being an instance of the machine learning model having the adapted parameters.