Device for hyper-dimensional computing tasks

Patent metadata

Field	Value
Publication number	US-11574209-B2
Application number	US-201916426572-A
Country	US
Kind code	B2
Filing date	May 30, 2019
Priority date	May 30, 2019
Publication date	Feb 7, 2023
Grant date	Feb 7, 2023

Abstract

A system for hyper-dimensional computing for inference tasks may be provided. The device comprises an item memory for storing hyper-dimensional item vectors, a query transformation unit connected to the item memory, the query transformation unit being adapted for forming a hyper-dimensional query vector from a query input and hyper-dimensional base vectors stored in the item memory, and an associative memory adapted for storing a plurality of hyper-dimensional profile vectors and for determining a distance between the hyper-dimensional query vector and the plurality of hyper-dimensional profile vectors, wherein the item memory and the associative memory are adapted for in-memory computing using memristive devices.

First claim

What is claimed is:

1. A device for hyper-dimensional computing for inference tasks, the device comprising
an item memory for storing hyper-dimensional item vectors,
query transformation unit connected to the item memory, the query transformation unit being adapted for forming a hyper-dimensional query vector from a query input and hyper-dimensional base vectors stored in the item memory, and
an associative memory adapted for storing a plurality of hyper-dimensional profile vectors and adapted for determining a distance between the hyper-dimensional query vector and the plurality of stored hyper-dimensional profile vectors,
wherein the item memory and the associative memory are adapted for in-memory computing using memristive devices.

2. The device according to claim 1 , wherein the in-memory computing is adapted to perform operations for inference tasks without altering the state of the memristive devices.

3. The device according to claim 1 , wherein the memristive devices are selected out of the group comprising, phase-change memory devices, ionic thin-film memristive devices, spin-based memristive devices, and magnetic memristive devices.

4. The device according to claim 1 , wherein the item memory comprises a crossbar array of memristive devices having as many memristive storage elements per row as the hyper-dimensional item vector has dimensions d.

5. The device according to claim 1 , wherein the associative memory comprises a crossbar array of memristive devices, and each profile hyper-dimensional vector is partitioned into equally sized (size p) profile sub-hyper-dimensional vectors, such that each profile sub-hyper-dimensional vector is encodable as binary conductance states into the memristive devices of one of a plurality of rows comprised in the associative memory.

6. The device according to claim 5 , also comprising:
a measurement unit adapted for measuring sub-hyper-dimensional vector currents resulting from applying query sub-hyper-dimensional vectors to corresponding profile sub-hyper-dimensional vectors, wherein the query sub-hyper-dimensional vectors are partitions of an equally partitioned (size p) query hyper-dimensional vector, and
an adder unit adapted for adding the respective sub-hyper-dimensional vector currents resulting in the distance between the query hyper-dimensional vector and the profile hyper-dimensional vector.

7. The device according to claim 1 :
wherein the item memory and the query transformation unit are portions of an encoder unit, and
where during training the hyper-dimensional profile vector per class to be predicted is determinable by binding shifted versions of selected item memory vectors to generate binary n-gram hyper-dimensional vectors, whereby intermediate results of the binding operation are storable in minterm buffers, by bundling together the n-gram hyper-dimensional vectors in a sum hyper-dimensional vector by summing the respective n-gram hyper-dimensional vector elements, and by applying a threshold on the elements of the sum hyper-dimensional vector to determine binary hyper-dimensional profile vector elements.

8. The device according to claim 1 , wherein the item memory comprises two symmetrical crossbar arrays, one being adapted for storing the hyper-dimensional item vector and the other one being adapted for storing a complement of the hyper-dimensional item vector.

9. The device according to claim 1 , wherein an output unit of the associative memory comprises a plurality of multiplexers, a summing buffer and a winner-takes-it-all (WTA) unit for a determination of a result of the query as a class output of an inference task.

10. The device according to claim 1 , wherein the distance between the hyper-dimensional query vector and a plurality of hyper-dimensional profile vectors in the associative memory is determined by the Hamming distance or by the dot-product between the respective vectors.

11. The device according to claim 1 , wherein the query transformation unit comprises for each dimension of the hyper-dimensional vectors a minterm buffer, a sum hyper-dimensional vector buffer, and a threshold unit.

12. The device according to claim 1 , wherein a controller, as a portion of the encoder, is adapted for issuing control signals orchestrating a data movement from a query symbol sequence to a query hyper-dimensional vector at the output of the encoder according to a configuration pertaining to a query symbol sequence received via a configuration interface.

13. The device according to claim 1 , wherein a second controller as a portion of an associative memory search module is adapted for iterating through each partition of the query sub-hyper-dimensional vector and activating corresponding partitions in the associative memory to obtain partial distances.

14. A method for hyper-dimensional computing for inference tasks, the method comprising:
storing hyper-dimensional item vectors in an item memory,
forming a hyper-dimensional query vector from a query input and hyper-dimensional base vectors stored in the item memory, and
determining a distance between the hyper-dimensional query vector and a plurality of hyper-dimensional profile vectors,
such that in-memory computing using memristive devices is performed during the storing, forming and determining.

15. The method according to claim 14 , wherein the memristive devices are selected out of the group comprising, phase-change memory devices, ionic thin-film memristive devices, spin-based memristive devices, and magnetic memristive devices.

16. The method according to claim 14 , wherein the item memory comprises a crossbar array of memristive devices having as many memristive storage elements per row as the hyper-dimensional item vector has dimensions d.

17. The method according to claim 14 , also comprising:
partitioning an associative memory comprising a crossbar array of memristive devices, and
partitioning each profile hyper-dimensional vector into equally sized (size p) profile sub-hyper-dimensional vectors, and
encoding each profile sub-hyper-dimensional vector as binary conductance states into the memristive devices of one of a plurality of rows comprised in the associative memory.

18. The method according to claim 17 , also comprising:
measuring sub-hyper-dimensional vector currents resulting from applying query sub-hyper-dimensional vectors to corresponding profile sub-hyper-dimensional vectors, wherein the query sub-hyper-dimensional vectors are partitions of an equally partitioned (size p) query hyper-dimensional vector, and
adding the respective sub-hyper-dimensional vector currents resulting in the distance between the query hyper-dimensional vector and the profile hyper-dimensional vector.

19. The method according to claim 14 , also comprising:
determining during training the hyper-dimensional profile vector per class to be predicted by binding shifted versions of selected item memory vectors to generate binary n-gram hyper-dimensional vectors, whereby intermediate results of the binding operation are stored in minterm buffers, by bundling together the n-gram hyper-dimensional vectors in a sum hyper-dimensional vector by summing the respective n-gram hyper-dimensional vector elements, and by applying a threshold on the elements of the sum hyper-dimensional vector to determine binary hyper-dimensional profile vector elements.

20. The method according to claim 14 , wherein the item memory comprises two symmetrical crossbar arrays, one being adapted for storing the hyper-dimensional item

Assignees

• IBM
• Eth Zuerich
• Eth Zurich Eidgenoessische Technische Hochschule Zurich

Inventors

• Karunaratne Kumudu Geethan
• Le Gallo-Bourdeau Manuel
• Cherubini Giovanni
• Sebastian Abu
• Rahimi Abbas
• Benini Luca

Classifications

• G06F9/4806

  Task transfer initiation or dispatching · CPC title

• G06G7/12

  Arrangements for performing computing operations, e.g. {operational} amplifiers specially adapted therefor · CPC title

• G06N5/04 Primary

  Inference or reasoning models · CPC title

• G06F40/00

  Handling natural language data (speech analysis or synthesis, speech recognition G10L) · CPC title

• G06F17/16

  Matrix or vector computation {, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization (matrix transposition G06F7/78)} · CPC title