Back to results
US 11,379,558 B2Granted
System enhancement methodology via matrix multiplication efficiency speedup using sparse basis approach
AssigneeNational Cheng Kung University
InventorsGwo Giun Lee, Shih-Yu Chen
Life of this patent
Estimating the term…
Abstract
The present invention relates to computing-implemented method and system that improves matrix multiplication efficiency, especially to method and system optimizing matrix multiplication using sparse basis approach. Matrices to be multiplied are organized into specially ordered vectors with zero values, facilitates speed up during linear combination computation or synthesis process.
Claims5
A computer-implement method of machine learning executed by one or more processors for feature extraction and classification, the method comprising acts of:
- identifying a first matrix and an intermediate matrix to be multiplied for producing a sparse basis matrix, wherein the first matrix is a matrix of input data, and the intermediate matrix is a filter used in the machine learning;
- applying the Gauss-Jordan Elimination (GJE) to the intermediate matrix to obtain a second matrix in reduced row echelon form (RREF), wherein the second matrix includes:
- the first non-zero term in the non-zero row being equal to 1, and being defined as a pivot;
- if a column contains a pivot, then the other entries in the column being zero;
- the location of the pivot in each non-zero row being from left to right and from the first row to last row;
- and
- zero rows being at the bottom of the second matrix;
- removing all zero rows from the second matrix for producing the sparse basis matrix;
- and
- determining an output matrix by multiplying a third matrix based on the sparse basis matrix, wherein the third matrix is a matrix denoting the coefficients required for a linear combination for the sparse basis matrix.
7 more elements
A system comprising:
- a memory for storing matrix data;
- and
- one or more processors being configured to perform matrix multiplication instruction to the matrix data stored in the memory using sparse basis approach in a convolutional neural network (CNN), wherein the operation of the processor comprising acts of
- identifying a first matrix and an intermediate matrix to be multiplied for producing a sparse basis matrix, wherein the first matrix is a matrix of input data, and the intermediate matrix is a filter of the CNN;
- applying the Gauss-Jordan Elimination (GJE) to the intermediate matrix to obtain a second matrix in reduced row echelon form (RREF);
- and
- determining an output matrix by multiplying a third matrix based on the sparse basis matrix, wherein the third matrix is a matrix denoting the coefficients required for a linear combination for the sparse basis matrix.
4 more elements
Dependencies are inferred from the claim text.