Invention Grant
- Patent Title: Parallel decoding using autoregressive machine learning models
-
Application No.: US16417190Application Date: 2019-05-20
-
Publication No.: US10521701B2Publication Date: 2019-12-31
- Inventor: Noam M. Shazeer , Jakob D. Uszkoreit , Mitchell Thomas Stern
- Applicant: Google LLC
- Applicant Address: US CA Mountain View
- Assignee: Google LLC
- Current Assignee: Google LLC
- Current Assignee Address: US CA Mountain View
- Agency: Fish & Richardson P.C.
- Main IPC: G06F15/18
- IPC: G06F15/18 ; G06K9/62 ; G06N20/00 ; G06N7/00

Abstract:
Methods, systems, and apparatus, including computer programs encoded on computer storage media, for performing parallel generation of output from an autoregressive sequence to sequence model. In one aspect, a blockwise parallel decoding method takes advantage of the fact that some architectures can score sequences in sublinear time. By generating predictions for multiple time steps at once then backing off to a longest prefix validated by the scoring model, the methods can substantially improve the speed of greedy decoding without compromising performance.
Public/Granted literature
- US20190354812A1 PARALLEL DECODING USING AUTOREGRESSIVE MACHINE LEARNING MODELS Public/Granted day:2019-11-21
Information query