公开(公告)号：US20210383199A1

公开(公告)日：2021-12-09

申请号：US16927018

申请日：2020-07-13

Applicant: Google LLC

Inventor： Dirk Weissenborn ,  Jakob Uszkoreit ,  Thomas Unterthiner ,  Aravindh Mahendran ,  Francesco Locatello ,  Thomas Kipf ,  Georg Heigold ,  Alexey Dosovitskiy

IPC: G06N3/063 ,  G06N3/08 ,  G06F17/16

Abstract: A method involves receiving a perceptual representation including a plurality of feature vectors, and initializing a plurality of slot vectors represented by a neural network memory unit. Each respective slot vector is configured to represent a corresponding entity in the perceptual representation. The method also involves determining an attention matrix based on a product of the plurality of feature vectors transformed by a key function and the plurality of slot vectors transformed by a query function. Each respective value of a plurality of values along each respective dimension of the attention matrix is normalized with respect to the plurality of values. The method additionally involves determining an update matrix based on the plurality of feature vectors transformed by a value function and the attention matrix, and updating the plurality of slot vectors based on the update matrix by way of the neural network memory unit.