Blockwise Hypervector Factorization for Faster Concept Decoding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for decoding hypervectors are resource-intensive due to the need to test every combination of factors, and sparse hypervectors complicate accurate factorization.
Innovation Solution
A resonator network system processes hypervectors blockwise, using superposition and clean-up memory strategies to efficiently search for factorizations without testing all combinations, and employs blockwise binding and unbinding operations to reduce computational complexity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If all possible combinations of factors are tested to decode hypervectors, then factorization accuracy is improved, but computational resources and time consumption increase significantly
Solution Approach 1:
The patent divides the high-dimensional hypervector into multiple blocks and processes each block separately through the resonator network. This segmentation allows the system to avoid testing all possible combinations of factors while maintaining factorization accuracy, as each block can be decoded independently and efficiently
Solution Approach 2:
The resonator network performs preliminary binding operations between query blocks and candidate factor blocks before full factorization. This preliminary action filters out incompatible factors early in the process, reducing the computational burden of testing all possible combinations while preserving accurate factorization results
2Device complexity
If dense hypervectors are used for factorization, then computational operations are simpler, but memory requirements and processing overhead increase
Solution Approach 1:
The patent employs sparse hypervectors where most elements are zero and only specific positions contain meaningful values. This local quality approach allows the resonator network to perform binding operations efficiently by focusing only on non-zero elements, reducing both computational complexity and memory requirements compared to dense hypervector representations
3Reliability
If iterative search is performed over all factor combinations, then complete factorization is achieved, but time consumption increases
Solution Approach 1:
The patent segments the factorization problem into independent block-level subproblems. The resonator network processes each block separately through iterative refinement, finding local optima that collectively achieve complete factorization. This segmentation dramatically reduces search time compared to exhaustive search over all possible factor combinations while maintaining factorization completeness
Solution Approach 2:
The resonator network employs periodic iterative refinement where estimates of factor blocks are updated in cycles. Each iteration refines the current estimates by performing binding operations and comparing results, gradually converging to the correct factorization. This periodic action achieves complete factorization much faster than exhaustive search by focusing computational effort on promising candidates
Data Source
AI summary
Embodiments are disclosed for a method. The method includes determining a granularity of hypervectors. The method also includes receiving an input hypervector representing a data structure. Additionally, the method includes performing an iterative process to factorize the input hypervector into individual hypervectors representing the cognitive concepts. The iterative process includes, for each concept: determining an unbound version of a hypervector representing the concept by a blockwise unbinding operation between the input hypervector and estimate hypervectors of other concepts. The iterative process further includes determining a similarity vector indicating a similarity of the unbound version of the hypervector with each candidate code hypervector of the concept. Additionally, the iterative process includes generating an estimate of a hypervector representing the concept by a linear combination of the candidate code hypervectors, and weights of the similarity vector.