Granted Patent B2

US 12,572,789 · App. 17/809,044 · Granted Mar 10, 2026

Inventors: Michael Andreas Hersche (Zurich, CH); Abu Sebastian (Adliswil, CH); Abbas Rahimi (Rüschlikon, CH)

Assignee: International Business Machines Corporation

G06N3/065

Abstract

Claims (87)

1 . A system comprising:

a computer processing circuit; and

a computer-readable storage medium storing instructions, which, when executed by the computer processing circuit, are configured to cause the computer processing circuit to perform a method comprising:

training an encoder comprising a feed-forward neural network to produce a hypervector s as a compound hypervector;

utilizing the encoder to represent data structures in a vector space, the vector space being defined by a set of codebooks that encode a set of cognitive concepts, wherein each of the set of codebooks comprise a plurality of candidate code hypervectors of dimension D representing a plurality of items of a concept of the set of cognitive concepts;

determining a granularity of a plurality of hypervectors so that each of the plurality of hypervectors includes a set of S blocks, each block having size L, wherein D=S×L;

receiving an input hypervector representing a data structure; and

performing an iterative process in order to factorize the input hypervector into a plurality of individual hypervectors representing the set of cognitive concepts, the iterative process comprising, for each concept of the set of cognitive concepts:

determining an unbound version of a hypervector representing the concept by a blockwise unbinding operation between the input hypervector and a plurality of estimate hypervectors of a plurality of other cognitive concepts;

determining a similarity vector indicating a similarity of the unbound version of the hypervector with each candidate code hypervector of the concept; and

generating an estimate of the hypervector representing the concept by a linear combination of the candidate code hypervectors, and a plurality of weights of the similarity vector.

2 . The system of claim 1 , wherein determining the similarity vector includes sparsifying the similarity vector before generating the estimate of the hypervector based on the sparsified similarity vector.

3 . The system of claim 2 , wherein determining the similarity vector includes adding a random noise vector to the similarity vector.

4 . The system of claim 2 , wherein sparsifying the similarity vector includes activating a portion of elements of the similarity vector, and deactivating by setting to a defined value a remaining portion of elements of the similarity vector.

5 . The system of claim 4 , wherein the activated elements of the similarity vector are a top j elements of the similarity vectors, wherein j is a configurable parameter that is smaller than a total number of elements of the similarity vector by a defined number.

6 . The system of claim 4 , wherein the activated elements of the similarity vector are the elements of similarity vectors having absolute values higher than a defined threshold.

7 . The system of claim 6 , the defined threshold being a mean of absolute values of elements of the similarity vector.

8 . The system of claim 4 , wherein sparsifying the similarity vector further comprises:

determining a maximum value of the activated vector;

comparing the maximum value with a threshold; if the maximum value exceeds the threshold, maintaining the maximum value; and

setting remaining elements of the activated vector to zero.

9 . The system of claim 1 , wherein generating the estimate of the hypervector comprises:

mapping the similarity vector to an S-dimensional vector such that each element of the S-dimensional vector indicates one candidate code hypervector of a codebook of the concept; and

building a new hypervector such that an i th block of the new hypervector is an i th block of a code hypervector indicated in an i th element of the S-dimensional vector.

10 . The system of claim 9 , the method further comprising, for every i th element of an S-dimensional vector s k ,

generating a value v, and determining

s

k

(

i

)

=

{

0

⁢

if

⁢

v

≤

∂

0

...

m

⁢

if

∂

m

-

1

<

v

≤

∂

m

where v is a randomly generated value or a deterministic value, v=i/S, where t=cumsum(a)/sum(a) is a step function, where a is the similarity vector, where cumsum(a) returns a vector containing a cumulative sum of elements of vector a, and sum(a) returns a sum of elements of vector a, and ∂ m refers to an m th element of ∂.

11 . The system of claim 1 , wherein determining the unbound version of the hypervector comprises:

a linear combination of the candidate code hypervectors, with weights given by the similarity vector;

an addition of a noise vector; and

an application of a selection function.

12 . The system of claim 1 , wherein determining the unbound version of the hypervector comprises:

a linear combination of the candidate code hypervectors, with weights given by the similarity vector; and

an application of a selection function.

13 . The system of claim 12 , wherein the selection function is applied blockwise.

14 . The system of claim 12 , wherein the selection function is a randomized argmax function.

15 . The system of claim 1 , further comprising:

responsive to a convergence criterion being fulfilled, stopping the iterative process.

16 . The system of claim 1 , wherein the hypervector includes binary values {0, 1} D and having a sparsity smaller than a sparsity threshold, wherein each block of the set of S blocks comprises a single non-zero value.

17 . A method comprising:

training an encoder comprising a feed-forward neural network to produce a hypervector s as a compound hypervector;

utilizing the encoder to represent data structures in a vector space, the vector space being defined by a set of codebooks that encode a set of cognitive concepts, wherein each of the set of codebooks comprise a plurality of candidate code hypervectors of dimension D representing a plurality of items of a concept of the set of cognitive concepts;

determining a granularity of a plurality of hypervectors so that each of the plurality of hypervectors includes a set of S blocks, each block having size L, wherein D=S×L;

receiving an input hypervector representing a data structure; and

performing an iterative process in order to factorize the input hypervector into a plurality of individual hypervectors representing the set of cognitive concepts, the iterative process comprising, for each concept of the set of cognitive concepts:

determining an unbound version of a hypervector representing the concept by a blockwise unbinding operation between the input hypervector and a plurality of estimate hypervectors of a plurality of other cognitive concepts;

determining a similarity vector indicating a similarity of the unbound version of the hypervector with each candidate code hypervector of the concept; and

generating an estimate of the hypervector representing the concept by a linear combination of the candidate code hypervectors, and a plurality of weights of the similarity vector.

18 . The method of claim 17 , wherein determining the similarity vector includes sparsifying the similarity vector before generating the estimate of the hypervector based on the sparsified similarity vector.

19 . A computer program product including program instructions stored on a computer readable storage medium, the program instructions executable by a processor to cause the processor to perform a method comprising:

training an encoder comprising a feed-forward neural network to produce a hypervector s as a compound hypervector;

utilizing the encoder to represent data structures in a vector space, the vector space being defined by a set of codebooks that encode a set of cognitive concepts, wherein each of the set of codebooks comprise a plurality of candidate code hypervectors of dimension D representing a plurality of items of a concept of the set of cognitive concepts;

determining a granularity of a plurality of hypervectors so that each of the plurality of hypervectors includes a set of S blocks, each block having size L, wherein D=S×L;

receiving an input hypervector representing a data structure; and

performing an iterative process in order to factorize the input hypervector into a plurality of individual hypervectors representing the set of cognitive concepts, the iterative process comprising, for each concept of the set of cognitive concepts:

determining an unbound version of a hypervector representing the concept by a blockwise unbinding operation between the input hypervector and a plurality of estimate hypervectors of a plurality of other cognitive concepts;

determining a similarity vector indicating a similarity of the unbound version of the hypervector with each candidate code hypervector of the concept; and

generating an estimate of the hypervector representing the concept by a linear combination of the candidate code hypervectors, and a plurality of weights of the similarity vector.

20 . The computer program product of claim 19 , the method further including sparsifying the similarity vector before generating the estimate of the hypervector based on the sparsified similarity vector.

Assignments (1)

Continuity (1)

References Cited (33)