Dimension Reduction Apparatus Using Segmented Partial Vectors
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional dimension reduction methods for high-dimensional feature vectors in pattern recognition, such as image or voice recognition, require large memory capacity and time due to the size of the Dout×Din matrix, and intermittent memory access slows down calculations.
Innovation Solution
A dimension reduction apparatus that generates multiple partial vectors and applies sequential mappings to reduce dimensions, using a combination of first and second mappings with parameter optimization, allowing for efficient storage and computation by reducing the number of parameters and enabling continuous memory access.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If principal component analysis is used to perform dimension reduction with a Dout×Din matrix, then dimension reduction accuracy is improved, but memory capacity requirement increases to extremely large values
Solution Approach 1:
The patent divides the high-dimensional feature vector into multiple partial vectors, each processed by corresponding partial mapping matrices. Instead of using one large Dout×Din matrix, the system segments both the input vector and the transformation matrix into manageable parts that can be processed independently and then combined, significantly reducing memory requirements while maintaining dimension reduction accuracy.
Solution Approach 2:
The patent transforms the problem from a single large matrix multiplication into multiple smaller matrix operations by introducing a segmentation dimension. The original Din-dimensional vector is divided into segments, and the transformation is performed through multiple smaller mapping matrices whose dimensions are much more manageable than the original Dout×Din matrix.
2Quantity of substance
If some elements of the Dout×Din matrix are set to nonzero elements to reduce memory capacity, then memory capacity requirement is reduced, but calculation speed decreases due to intermittent memory access
Solution Approach 1:
The patent segments the transformation process into multiple partial mappings, where each partial mapping operates on a subset of the input dimensions. This segmentation allows each partial mapping matrix to be stored in a compact form and accessed sequentially, avoiding the intermittent memory access patterns that occur with sparse matrices and thereby maintaining high calculation speed while reducing memory usage.
Solution Approach 2:
The patent ensures continuous memory access during calculation by organizing the partial mapping operations to access memory sequentially. Each partial mapping matrix is processed in a continuous manner without gaps, eliminating the intermittent access patterns that slow down sparse matrix operations, thus maintaining high productivity.
3Measurement precision
If a large Dout×Din matrix is used for dimension reduction, then recognition accuracy is improved, but calculation time increases to enormous proportions
Solution Approach 1:
The patent divides the single large matrix multiplication operation into multiple smaller partial mapping operations. Each partial mapping processes a segment of the input vector and produces a corresponding segment of the output. This segmentation reduces the computational complexity from O(Dout×Din) to a sum of smaller operations, significantly reducing calculation time while preserving the overall transformation accuracy needed for recognition.
Solution Approach 2:
The patent applies partial mapping operations to segments of the input vector rather than processing the entire high-dimensional vector at once. This partial action approach allows the system to perform dimension reduction in manageable steps, reducing the total calculation time required while still achieving the necessary recognition accuracy through the combination of partial results.
Data Source
AI summary
According to an embodiment, a dimension reduction apparatus includes an obtaining device to obtain a first vector made up of elements; a generator to generate first partial vectors each made up of one or more elements of the elements; a first mapping controller to transform each first partial vector into a second partial vector by reducing number of elements constituting the first partial vector using first mapping corresponding to the first partial vector, concatenate the second partial vectors, and generate a second vector by reducing number of dimensions of the first vector; and a second mapping controller to apply, to the second vector, second mapping which maps the second vector into a space having number of dimensions not larger than that of the second vector, and generate a third vector by reducing the number of dimensions of the second vector to not larger than that of the second vector.