Efficient look-up for vector symbolic architectures (vsa)

Patent metadata

Field	Value
Publication number	US-2025258826-A1
Application number	US-202418436481-A
Country	US
Kind code	A1
Filing date	Feb 8, 2024
Priority date	Feb 8, 2024
Publication date	Aug 14, 2025
Grant date	—

Abstract

A method for hyperdimensional computing to obtain an answer to a query includes encoding each of an N number of data points related to the query with a first high dimensional vector P and a second high dimensional vector H to generate an encoded N number of data points having a P vector component and an H vector component, processing the encoded N number of data points via a first sub-routine to generate an intermediate result, wherein the first sub-routine is responsive to Vector Symbolic Architecture (VSA) operations, processing the intermediate result via a second sub-routine to generate a final result, wherein the second sub-routine is responsive to Vector Symbolic Architecture (VSA) operations and conducting a similarity search of the P vector component of the final result to generate an answer to the query.

First claim

1 . A method for hyperdimensional computing to obtain an answer to a query, the method comprising:
encoding each of an N number of data points related to the query with a first high dimensional vector P and a second high dimensional vector H to generate an encoded N number of data points having a P vector component and an H vector component; processing the encoded N number of data points via a first sub-routine to generate an intermediate result, wherein the first sub-routine is responsive to Vector Symbolic Architecture (VSA) operations; processing the intermediate result via a second sub-routine to generate a final result, the second sub-routine including a linear search where, for each result j, a similarity (H j ) is calculated, and a largest similarity (H j ) is identified, where the similarity (H j ) is given by: H j =[Result/H j ], and wherein the second sub-routine is responsive to Vector Symbolic Architecture (VSA) operations; and conducting a similarity search of the P vector component of the final result to generate an answer to the query, wherein the similarity search of the P vector retrieves the similarity H without performing a full similarity search.

2 . The method of claim 1 , wherein the N number of data points are selected responsive to the query.

3 . The method of claim 1 , wherein the H vector component is larger than the P vector component.

4 . The method of claim 1 , wherein the first sub-routine includes performing a bundling/superposition operation on the P vector component and the H vector component to generate an N number of bundled data points within the P vector component and the H vector component which are bundled together.

5 . The method of claim 4 , wherein the first sub-routine further includes performing a binding operation on the N number of bundled data points to generate the intermediate result having an N number of bound data points within the P vector component and the H vector component.

6 . The method of claim 1 , wherein the second sub-routine includes performing a unbinding operation on the N number of bound data points to generate an N number of unbound data points within the P vector component and the H vector component.

7 . The method of claim 6 , wherein the second sub-routine further includes performing a permutation operation on the N number of unbound data points to generate the final result having an N number of permutated data points within the P vector component and the H vector component.

8 . A non-transitory computer readable medium storing instructions configured to cause a computer system to implement operations comprising:
encoding each of an N number of data points related to the query with a first high dimensional vector P and a second high dimensional vector H to generate an encoded N number of data points having a P vector component and an H vector component; processing the encoded N number of data points via a first sub-routine to generate an intermediate result, wherein the first sub-routine is responsive to Vector Symbolic Architecture (VSA) operations; processing the intermediate result via a second sub-routine to generate a final result, wherein the second sub-routine is responsive to Vector Symbolic Architecture (VSA) operations, the second sub-routine including a linear search where, for each result j, a similarity (H j ) is calculated, and a largest similarity (H j ) is identified, where the similarity (Hi) is given by: H j =[Result/H j ]; and conducting a similarity search of the P vector component of the final result to generate an answer to the query, wherein the similarity search of the P vector retrieves the similarity H without performing a full similarity search.

9 . The non-transitory computer readable medium of claim 8 , wherein the N number of data points are selected responsive to the query.

10 . The non-transitory computer readable medium of claim 8 , wherein the H vector component is larger than the P vector component.

11 . The non-transitory computer readable medium of claim 8 , wherein the first sub-routine includes performing a bundling/superposition operation on the P vector component and the H vector component to generate an N number of bundled data points within the P vector component and the H vector component which are bundled together.

12 . The non-transitory computer readable medium of claim 11 , wherein the first sub-routine further includes performing a binding operation on the N number of bundled data points to generate the intermediate result having an N number of bound data points within the P vector component and the H vector component.

13 . The non-transitory computer readable medium of claim 8 , wherein the second sub-routine includes performing a unbinding operation on the N number of bound data points to generate an N number of unbound data points within the P vector component and the H vector component.

14 . The non-transitory computer readable medium of claim 13 , wherein the second sub-routine further includes performing a permutation operation on the N number of unbound data points to generate the final result having an N number of permutated data points within the P vector component and the H vector component.

15 . A computer program product comprising a non-transitory computer readable storage medium having program instructions embodied therewith, the program instructions executable by a processor to cause the processor to perform operations for implementing a method for hyperdimensional computing to obtain an answer to a query, the method comprising:
encoding each of an N number of data points related to the query with a first high dimensional vector P and a second high dimensional vector H to generate an encoded N number of data points having a P vector component and an H vector component; processing the encoded N number of data points via a first sub-routine to generate an intermediate result, wherein the first sub-routine is responsive to Vector Symbolic Architecture (VSA) operations; processing the intermediate result via a second sub-routine to generate a final result, wherein the second sub-routine is responsive to Vector Symbolic Architecture (VSA) operations, the second sub-routine including a linear search where, for each result j, a similarity (H j ) is calculated, and a largest similarity (H j ) is identified, where the similarity (H j ) is given by: H j =[Result/H j ]; and conducting a similarity search of the P vector component of the final result to generate an answer to the query, wherein the similarity search of the P vector retrieves the similarity H without performing a full similarity search.

16 . The computer program product of claim 15 , wherein
the N number of data points are selected responsive to the query, and the H vector component is larger than the P vector component.

17 . The computer program product of claim 15 , wherein the first sub-routine includes performing a bundling/superposition operation on the P vector component and the H vector component to generate an N number of bundled data points within the P vector component and the H vector component which are bundled together.

18 . The computer program product of claim 17 , wherein the first sub-routine further includes performing a binding operation on the N number of bundled data points to generate the intermediate result having an N number of bound data points within the P vector component and the H vector component.

19 . The computer program product of claim 15 , wherein the second sub-routine includes performing a unbi

Assignees

• IBM

Inventors

• Gokmen Tayfun
• Kalantzis Vasileios
• Ram Parikshit
• Wu Chai Wah
• Clarkson Kenneth Lee
• Horesh Lior
• Ubaru Shashanka

Classifications

• G06F16/24561 Primary

  Intermediate data storage techniques for performance improvement · CPC title

• G06F16/2237

  Vectors, bitmaps or matrices · CPC title

• G06F16/2264 Primary

  Multidimensional index structures · CPC title