Sparse Convolution Fetch Control in Neural Network Processors
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural network systems face challenges in efficiently processing complex input data in real time, necessitating improved apparatuses and methods for neural network operations.
Innovation Solution
A neural network processor with a fetch controller and data arithmetic circuit that selectively convolute input features and weights based on non-zero values, and a controller that groups weight maps and neural network processors to optimize operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If all input features and weights are convoluted in neural network processing, then complete data processing is achieved, but operation time and computational load increase significantly
Solution Approach 1:
The patent extracts and identifies only the non-zero input features and weights from the complete data sets, separating meaningful data from zero-value data. The fetch controller detects non-zero values and extracts only those specific features and weights for convolution, eliminating unnecessary zero-value computations while preserving complete processing of meaningful data.
Solution Approach 2:
The patent applies partial action by performing convolution operations only on the subset of non-zero input features and weights rather than processing all features and weights completely. This selective partial processing maintains data processing completeness for meaningful values while reducing overall operation time by omitting redundant zero-value operations.
2Loss of time
If selective convolution of non-zero features and weights is implemented, then operation time is reduced, but processing complexity increases due to detection requirements
Solution Approach 1:
The patent applies preliminary action by having the fetch controller detect and identify non-zero input features and weights before the convolution operation begins. This pre-detection phase prepares the data by marking or separating non-zero values, so that the subsequent convolution process only needs to process the already-identified meaningful data without needing to evaluate zero-values, thereby reducing operation time while managing complexity through staged processing.
3Productivity
If zero-value computations are omitted in convolution, then computational efficiency improves, but data processing completeness may be compromised
Solution Approach 1:
The patent extracts only the non-zero input features and weights for convolution operations, separating meaningful computational data from zero-value data. This extraction ensures that zero-value computations are omitted to improve computational efficiency while the complete set of non-zero features and weights is still processed to maintain data processing completeness.
Solution Approach 2:
The patent changes the parameter selection criteria for convolution from processing all features and weights to processing only those with non-zero values. This parameter change in the convolution operation - selecting based on non-zero value status rather than complete inclusion - improves computational efficiency by eliminating redundant zero-value operations while maintaining completeness of meaningful data processing.
Data Source
AI summary
A neural network processor and method include a fetch controller configured to receive input feature information, indicating whether each of a plurality of input features of an input feature map includes a non-zero value, and weight information, indicating whether each of a plurality of weights of a weight map includes a non-zero value, and configured to determine input features and weights to be convoluted, from among the plurality of input features and the plurality of weights, based on the input feature information and the weight information. The neural network processor and method also include a data arithmetic circuit configured to convolute the determined weights and input features to generate an output feature map.