Matrix Convolution Processing for Sparse 2D and 3D AI Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for performing convolution operations in artificial intelligence systems, particularly those using convolutional neural networks, face inefficiencies due to the processing of data sparsity in one-dimensional arrays, failing to optimize operations when dealing with two-dimensional or three-dimensional data where many elements are zeros, leading to unnecessary multiplications.
Innovation Solution
An electronic apparatus with a processor featuring a matrix formation of processing elements that compares and identifies non-zero values in object and kernel data, allowing only those elements to perform convolution operations, thereby reducing operation cycles by omitting unnecessary multiplications.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If conventional methods process data sparsity in one-dimensional arrays, then the processing structure is simple, but the operation speed is slow and unnecessary multiplications are performed
Solution Approach 1:
The patent transforms the processing structure from one-dimensional arrays to two-dimensional matrix formation of processing elements. This dimensional change enables parallel processing of multiple data elements simultaneously, significantly improving operation speed while efficiently handling data sparsity in 2D and 3D convolution operations
2Productivity
If all elements are processed in convolution operations, then complete computation is performed, but unnecessary multiplications increase operation cycles
Solution Approach 1:
The patent extracts and identifies non-zero elements from the input data using logic circuits before performing multiplication operations. By taking out only the necessary non-zero elements for computation and excluding zero elements, the system reduces unnecessary multiplications and decreases the total number of operation cycles required
Solution Approach 2:
The patent performs preliminary identification of non-zero elements through logic circuits before the actual multiplication operation. This preliminary action of filtering and selecting only relevant data elements ahead of time optimizes the subsequent computation process and improves overall operation efficiency
Data Source
AI summary
An electronic apparatus is provided. The electronic apparatus includes a storage storing object data and kernel data, and a processor including a plurality of processing elements arranged in a matrix formation, wherein the processor is configured to input corresponding first elements among a plurality of first elements included in the object data into processing elements arranged in a first row among the plurality of processing elements, and input a plurality of second elements included in the kernel data sequentially into the processing elements arranged in the first row to perform operations between the corresponding first elements and the plurality of second elements, to identify a depth in which a first element and a second element have a non-zero value, and to input the first element and the second element corresponding to the identified depth into a calculator included in each of the processing elements arranged in the first row to perform a convolution operation.