Granted Patent B2

US 12,306,906 · App. 17/525,908 · Granted May 20, 2025

Inventors: Mohsen Fayyaz (Bonn, DE); Soroush Abbasi Koohpayegani (Baltimore, MD); Eric Chris Wolfgang Sommerlade (Oxford, GB); Hamidreza Vaezi Joze (Redmond, WA)

Assignee: Microsoft Technology Licensing, LLC

G06F18/2113G06F18/24G06N3/04

Abstract

Claims (43)

1. A computer-implemented method for processing a data item, comprising:

obtaining plural item tokens that represent the data item;

obtaining a classification token; and

converting the item tokens and the classification token into embedding vectors;

in an attention operation, using a transformer neural network for: generating original attention information based on the embedding vectors, the original attention information having a plurality of attention values, each attention value describing an importance that a particular token plays in an interpretation of another particular token; generating score information based on attention values in the original attention information that pertain to the classification token; and generating modified attention information by removing attention values from the original attention information, as guided by a sampling operation that is performed based on the score information; and

performing subsequent operations in the transformer neural network based on the modified attention information,

the subsequent operations performing fewer operations by using the modified attention information rather than the original attention information.

2. The computer-implemented method of claim 1 , wherein the data item is an image, and wherein the item tokens are image tokens that represent respective parts of the image.

3. The computer-implemented method of claim 1 , wherein the item tokens represent parts of a feature map produced by performing a convolution operation on the data item.

4. The computer-implemented method of claim 1 , wherein the transformer neural network includes plural encoder blocks, wherein a particular encoder block of the plural encoder blocks performs the attention operation.

5. The computer-implemented method of claim 1 , wherein the transformer neural network includes plural encoder blocks, and wherein two or more of the plural encoder blocks perform instances of the attention operation.

6. The computer-implemented method of claim 1 , wherein the transformer neural network outputs transformer output information, and wherein the method further comprises classifying item content in the data item based on the transformer output information.

7. The computer-implemented method of claim 1 , wherein the sampling operation comprises choosing no more than a prescribed number of item tokens based on the score information, the prescribed number being less than a total number of item tokens.

8. The computer-implemented method of claim 1 , wherein the sampling operation comprises choosing item tokens using inverse transform sampling based on the score information.

9. The computer-implemented method of claim 1 , wherein the sampling operation comprises:

ordering individual scores of the score information by score magnitudes, to provide ordered score information;

generating a cumulative distribution function based on the ordered score information; and

non-uniformly selecting item tokens using the cumulative distribution function.

10. The computer-implemented method of claim 1 , wherein the original attention information is expressed as an array of the attention values, and wherein said generating modified attention information removes one or more groups of attention values in the original attention information.

11. The computer-implemented method of claim 1 , wherein said generating score information and said generating modified attention information are functions of the transformer neural network that do not have their own machine-trained weights.

12. A transformer neural network, comprising:

hardware logic circuitry configured to perform operations based on a set of machine-trained weights, the operations including:

receiving embedding vectors that represent a plurality of item tokens generated based on a data item, and a classification token;

in an attention operation: generating original attention information based on the embedding vectors, the original attention information having a plurality of attention values, each attention value describing an importance that a particular token plays in an interpretation of another particular token; generating score information based on attention values in the original attention information that pertain to the classification token; and generating modified attention information by removing attention values from the original attention information, as guided by a sampling operation that is performed based on the score information; and

performing subsequent operations based on the modified attention information,

the subsequent operations performing fewer operations by using the modified attention information rather than the original attention information.

13. The transformer neural network of claim 12 , wherein the transformer neural network includes plural encoder blocks, wherein a particular encoder block of the plural encoder blocks performs the attention operation.

14. The transformer neural network of claim 12 , wherein the transformer neural network includes plural encoder blocks, and wherein two or more of the plural encoder blocks perform instances of the attention operation.

15. The transformer neural network of claim 12 , wherein the sampling operation comprises choosing no more than a prescribed number of item tokens based on the score information, the prescribed number being less than a total number of item tokens.

16. The transformer neural network claim 12 , wherein the sampling operation comprises choosing item tokens using inverse transform sampling based on the score information.

17. The transformer neural network of claim 12 , wherein the sampling operation comprises:

ordering individual scores of the score information by score magnitudes, to provide ordered score information;

generating a cumulative distribution function based on the ordered score information; and

non-uniformly selecting item tokens using the cumulative distribution function.

18. The transformer neural network of claim 12 , wherein the original attention information is expressed as an array of attention values, and wherein said generating modified attention information produces the modified attention information by removing one or more groups of attention values in the original attention information.

19. The transformer neural network of claim 12 , wherein said generating score information and said generating modified attention information are functions of the transformer neural network that do not have their own machine-trained weights.

20. A non-transitory computer-readable storage medium for storing computer-readable instructions, the computer-readable instructions, when executed by one or more hardware processors, performing a method that comprises:

receiving embedding vectors that represent a plurality of item tokens generated based on a data item, and a classification token;

generating original attention information based on the embedding vectors, the original attention information having a plurality of attention values, each attention value describing an importance that a particular token plays in an interpretation of another particular token;

generating score information based on attention values in the original attention information that pertain to the classification token;

generating modified attention information by removing attention values from the original attention information, as guided by a sampling operation that is performed based on the score information;

performing subsequent operations in the transformer neural network based on the modified attention information, to produce output information,

the method being performed on a basis of machine-trained weights, the machine-trained weights governing said generating original attention information, but not describing functions performed by said generating score information and said generating modified attention information.

Assignments (1)

Continuity (1)

References Cited (144)