Hardware Selector for Sparsified Convolutional Neural Network Operations
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Convolutional neural networks are operation-intensive, and existing methods fail to efficiently perform convolution operations, especially when weight parameters are sparsified, leading to increased computational and storage inefficiencies.
Innovation Solution
An apparatus comprising a selector for nonzero elements in weight parameters and input feature data, coupled with a calculator for efficient multiplication and accumulation, allowing for the performance of convolution operations in sparsified convolutional neural networks through hardware implementation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If weight parameters are sparsified to reduce operations, then computational efficiency is improved, but hardware implementation complexity increases due to selective processing requirements
Solution Approach 1:
The weight parameters are segmented into nonzero elements and zero elements. The selector apparatus segments the weight parameters and input feature data to extract only nonzero elements for processing, while the calculator processes only the segmented nonzero portions, avoiding computation on zero elements and thereby improving computational efficiency without requiring complete restructuring of the hardware
Solution Approach 2:
The selector extracts only the nonzero elements from weight parameters and corresponding input feature data, separating them from zero elements. This extraction mechanism allows the calculator to process only relevant data, reducing unnecessary computations while maintaining a relatively simple hardware structure through targeted data selection rather than complex processing logic
2Measurement precision
If all weight parameters are processed in convolution operations, then accuracy is maintained, but computational load increases significantly
Solution Approach 1:
Different quality processing is applied to different weight parameters: nonzero elements undergo full multiplication and accumulation operations to maintain accuracy, while zero elements are excluded from processing entirely. The selector identifies and routes only nonzero elements to the calculator, ensuring that computational resources are focused on elements that contribute to accuracy while eliminating wasteful computations on zero elements
Solution Approach 2:
Instead of processing all weight parameters uniformly, the system applies partial action by processing only the nonzero elements that contribute to the convolution result. This partial processing approach maintains recognition accuracy for the contributing elements while avoiding excessive computational load on elements that would produce zero contribution anyway
3Productivity
If zero value elements are skipped in convolution operations, then operation count is reduced, but data selection overhead increases
Solution Approach 1:
The selection of nonzero elements from weight parameters and corresponding input feature data is merged into a coordinated process. The selector apparatus combines the selection of weight nonzero elements with the selection of corresponding input data items, using the element positions as joint indexing keys. This merging reduces redundant selection operations and minimizes data selection overhead while maintaining reduced operation counts
Data Source
AI summary
Disclosed is an apparatus for performing a convolution operation in a convolutional neural network. The apparatus may comprise a selector for selecting one or more nonzero elements of a weight parameter, a selector for selecting a data item(s) corresponding to selected nonzero elements in input feature data, and a calculator unit for performing an operation. The apparatus may realize the convolution operation in a sparsified convolutional neural network efficiently through the hardware.