Relatedness Determination Using Binary Basis Vectors
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for determining relatedness between feature vectors and real vectors, such as in image and speech recognition, face high computational loads and memory consumption due to the use of floating-point representations, which are exacerbated by increasing feature vector dimensions.
Innovation Solution
The method involves decomposing real vectors into a linear sum of binary or ternary basis vectors and performing inner product calculations between these basis vectors and binarized feature vectors to speed up the relatedness determination process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If floating-point feature vectors are used for relatedness determination, then measurement precision is maintained, but computational load and memory consumption increase significantly
Solution Approach 1:
The patent changes the parameter representation from floating-point to binary/ternary discrete values. Real vectors are transformed into binary codes through hashing functions, and feature vectors are converted into binary representations. This parameter transformation maintains the essential information for relatedness determination while dramatically reducing computational complexity and memory requirements.
Solution Approach 2:
The patent creates a simplified copy of the real vector space using binary/ternary basis vectors. Instead of working with the original floating-point vectors directly, the system uses discrete approximations that capture the essential relationships. The inner product calculation between binary feature vectors and real vectors (decomposed into basis vectors) provides a computationally efficient proxy for the original complex calculation.
2Measurement precision
If feature vector dimension is increased to improve measurement precision, then relatedness determination accuracy improves, but calculation load and memory consumption increase linearly
Solution Approach 1:
The patent applies parameter changes by transforming high-dimensional floating-point vectors into binary/ternary representations. The hashing function converts real-valued features into discrete binary codes, and the decomposition of real vectors into basis vectors with discrete values reduces the effective dimensionality while maintaining discrimination capability. This reduces the complexity of vector operations significantly.
Solution Approach 2:
The patent segments the real vector into a linear combination of basis vectors with discrete values. By decomposing the real vector q into basis vectors {bi} with coefficients ci, the system breaks down the complex high-dimensional space into simpler components. This segmentation allows the inner product calculation to be performed more efficiently by working with the discrete basis representations rather than the full high-dimensional space.
3Productivity
If binary code conversion is applied to reduce computational load, then calculation speed improves, but the ability to handle real vector operations is limited
Solution Approach 1:
The patent introduces real vectors decomposed into basis vectors as an intermediary between the binary feature vectors and the original real vector space. The basis vectors {bi} serve as a bridge, allowing the system to perform operations in the simplified binary space while maintaining the ability to represent and manipulate real vector relationships through the decomposition q = Σci·bi.
Solution Approach 2:
The patent makes the binary/ternary basis vector representation universal by showing it can handle multiple operations: inner product calculations, similarity measurements, and real vector representations. The same discrete basis representation works for both the feature vectors and the real vectors, providing a unified approach that maintains versatility while improving computational efficiency.
Data Source
AI summary
A relatedness determination device includes: a feature vector acquisition portion that acquires a binarized feature vector; a basis vector acquisition portion that acquires a plurality of basis vectors obtained by decomposing a real vector into a linear sum of the basis vectors, which have a plurality of elements including only binary or ternary discrete values; and a vector operation portion that sequentially performs inner product calculation between the binarized feature vector and each of the basis vectors to determine relatedness between the real vector and the binarized feature vector.