Neural Network Convolution Redundancy Reduction via Base Planes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current neural network convolution operations are inefficient due to redundant calculations caused by low-bit precision weights, leading to increased computing resources and time consumption, especially in resource-limited environments.
Innovation Solution
The method involves generating base planes from available weight values of a weight kernel, determining target regions based on weight element offsets, and performing accumulation operations using shift and addition instead of multiplication, to reduce redundancy and optimize convolution operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If low-bit precision weights are used in convolution operations, then device complexity and computing resource consumption are reduced, but calculation accuracy and network performance deteriorate due to redundant operations
Solution Approach 1:
The patent applies preliminary action by pre-calculating base planes from the weight kernel before the actual convolution operation. The base planes are generated in advance using available weight values, and then reused during the convolution process through accumulation operations. This eliminates redundant calculations while maintaining accuracy, as the base planes are computed once and then efficiently accumulated with input feature maps.
2Measurement precision
If traditional multiplication operations are performed for each weight element, then calculation accuracy is maintained, but processing speed and efficiency deteriorate due to redundant multiplications
Solution Approach 1:
The patent merges multiple multiplication operations into a single accumulation operation. Instead of performing separate multiplication operations for each weight element, the method combines them by accumulating values from base planes that have been pre-computed. This merging of operations significantly reduces the number of computational steps while maintaining the same mathematical result, thereby improving processing speed without sacrificing accuracy.
Solution Approach 2:
The patent changes the computational parameter from multiplication to accumulation (addition). By transforming the operation type from multiplication to accumulation, the patent exploits the property that accumulation is computationally less expensive and can be more efficiently implemented in hardware. This parameter change maintains numerical accuracy while dramatically improving processing speed.
3Productivity
If redundancy reduction techniques are applied, then processing efficiency is improved, but device complexity increases due to additional base plane generation and management
Solution Approach 1:
The patent segments the weight kernel into base planes based on available weight values. By dividing the weight kernel into discrete base planes that can be independently generated and managed, the patent enables efficient reuse of these segments during convolution. This segmentation reduces overall complexity by organizing the computation into reusable units rather than requiring full recalculation for each operation.
Data Source
AI summary
A processor-implemented neural network layer convolution operation method includes: obtaining a first input plane of an input feature map and a first weight plane of a weight kernel; generating base planes, corresponding to an intermediate operation result of the first input plane, based on at least a portion of available weight values of the weight kernel; generating first accumulation data based on at least one plane corresponding to weight element values of the first weight plane among the first input plane and the base planes; and generating a first output plane of an output feature map based on the first accumulation data.