Title:
ATTENTION BASED SEQUENCE CONVERSION NEURAL NETWORK
Document Type and Number:
Japanese Patent JP2023052483
Kind Code:
A
Abstract:
To provide a system for converting a sequence using a neural network and a plurality of computer storage media.SOLUTION: The present invention is directed to a neural network system having one or a plurality of encoder neural networks 110 for receiving an input sequence and generating expressions obtained by encoding network inputs in the input sequence; and a decoder neural network 150 for receiving the encoded expressions to generate an output sequence. An encoder subnetwork 130 for receiving an encoder subnetwork input of each of a plurality of input positions to generate a subnetwork output for each of the plurality of input positions has an encoder self-attention sublayer 132 for receiving the subnetwork input for each of the plurality of input positions.SELECTED DRAWING: Figure 1
Inventors:
Nome M Shazir
Aidan Nicholas Gomez
Lucas Mieticslav Kaiser
Jacob Dee Utsukolite
Rion Owen Jones
Niki Jay Palmer
Ilya Polo Skin
Ashish Tech Vaswani
Aidan Nicholas Gomez
Lucas Mieticslav Kaiser
Jacob Dee Utsukolite
Rion Owen Jones
Niki Jay Palmer
Ilya Polo Skin
Ashish Tech Vaswani
Application Number:
JP2023006053A
Publication Date:
April 11, 2023
Filing Date:
January 18, 2023
Export Citation:
Assignee:
Google LLC
International Classes:
G06N3/0455
Attorney, Agent or Firm:
Murayama Yasuhiko
Shinya Mihiro
Tatsuhiko Abe
Shinya Mihiro
Tatsuhiko Abe