3D Convolution Kernel Segmentation for CNN Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current 2D-configured CNN processors face inefficiencies when performing 3D convolution operations due to the need for padding, which they are not equipped to handle efficiently, leading to suboptimal processing times and resource allocation.
Innovation Solution
The solution involves determining whether a convolution iteration is of a first or second type and replacing inefficient iterations with efficient ones by skipping calculations between kernel and padding segments, such as replacing kernel elements with zero-values and performing operations only with input data segments.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If 2D-configured CNN processors perform 3D convolution operations with padding, then object detection and classification can be achieved, but processing time increases and efficiency decreases
Solution Approach 1:
The patent segments the 3D convolution operation into distinct types (first type and second type iterations). The processor identifies and handles different segments of convolution operations differently - using standard processing for first type and optimized skipping for second type, thereby reducing overall processing time while maintaining 3D convolution capability
Solution Approach 2:
The patent transitions from traditional 2D convolution processing to 3D convolution processing by adding the depth dimension. This is achieved through virtual padding that creates a third dimension without requiring physical hardware changes, enabling 3D operations on 2D-configured processors
2Manufacturing precision
If virtual padding segments are included in convolution calculations, then 3D convolution structure is maintained, but computational efficiency decreases due to futile calculations
Solution Approach 1:
The patent extracts and removes the virtual padding segments from the computational process. By identifying second type iterations where only virtual padding is involved, the processor skips these futile calculations entirely, eliminating wasted computational effort while preserving the necessary 3D convolution structure through selective processing
Solution Approach 2:
The patent implements a skipping mechanism that allows the processor to rush through or bypass second type convolution iterations that involve only virtual padding segments. This skipping approach maintains computational accuracy for meaningful operations while eliminating time-wasting calculations on padding-only regions
Data Source
AI summary
A method for neural network convolution, the method may include receiving input data that is a 3D input data and comprises input data segments associated with different input data depth values; receiving a convolution kernel that is a 3D convolution kernel and comprises kernel segments associated with different kernel depth values; performing multiple 3D convolution iteration, wherein each of 3D convolution iteration comprises: determining whether the 3D convolution iteration is of a first type or of a second type; executing the 3D convolution iteration of the first type when determining that the 3D convolution iteration is of the first type; and executing the 3D convolution iteration of the second type when determining that the 3D convolution iteration is of the second type.