Shared-Operand Convolution Processing for Lower Computational Complexity
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural network processing systems face inefficiencies in performing convolution operations due to the large number of operations required, particularly in deep neural networks, which can lead to increased computational complexity and resource utilization.
Innovation Solution
A neural network apparatus that determines a shared operand, either a pixel value of an input feature map or a weight value of a kernel, to perform parallelized operations, allowing for efficient utilization of processing units and reducing computational overhead by implementing zero skipping when the shared operand is zero.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional convolution operations are performed in neural networks, then computational accuracy is maintained, but computational complexity and resource utilization increase significantly
Solution Approach 1:
The patent segments the convolution operation into two distinct phases: a first convolution operation that processes the entire input feature map to generate a first feature map, and a second convolution operation that processes only selected pixel values from the first feature map to generate the output feature map. This segmentation reduces the total number of computational operations while maintaining accuracy by strategically selecting which pixel values require further processing.
Solution Approach 2:
The patent extracts and identifies specific pixel values from the first feature map that meet predetermined conditions (such as having maximum absolute values or exceeding threshold values). By taking out only these significant pixel values for the second convolution operation, the system avoids performing computations on pixel values that would not contribute meaningfully to the final output, thereby reducing computational complexity while preserving accuracy.
2Reliability
If traditional convolution operations are performed in neural networks, then complete processing is achieved, but resource utilization increases
Solution Approach 1:
The patent applies local quality by differentiating the processing intensity applied to different regions of the feature map. Instead of uniformly processing all pixel values with the same computational resources, the system identifies and applies the second convolution operation only to local regions containing pixel values that meet specific criteria (such as having maximum absolute values or exceeding thresholds). This localized approach ensures processing completeness for critical regions while conserving resources in less critical regions.
Solution Approach 2:
The patent implements partial action by performing the second convolution operation on only a subset of pixel values rather than all pixel values. By applying the operation partially—to only those pixel values that satisfy predetermined conditions—the system achieves sufficient processing reliability for the most important data points while significantly reducing overall resource utilization compared to complete processing of all pixels.
3Speed
If all pixel values are processed in parallel, then processing speed increases, but hardware complexity increases
Solution Approach 1:
The patent implements dynamic processing by adaptively determining the number and configuration of processing units based on the actual data characteristics. Instead of using a fixed, static hardware architecture that processes all pixels in parallel, the system dynamically identifies how many pixel values meet the predetermined conditions and configures the appropriate number of processing units accordingly. This dynamic approach maintains high processing speed for significant data points while avoiding the hardware complexity of provisioning for maximum possible parallelism in all cases.
Data Source
AI summary
A neural network apparatus including one or more processors including a controller configured to determine a shared operand to be shared in parallelized operations as being either one of a pixel value among pixel values of an input feature map and a weight value among weight values of a kernel, based on either one or both of a feature of the input feature map and a feature of the kernel, and one or more processing units configured to perform the parallelized operations based on the determined shared operand.