Dilated Convolution Feature Map Splitting for Neural Network Efficiency
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Low-power neural network devices face challenges in efficiently processing large amounts of input data in real-time due to high computational requirements for complex operations, leading to inefficiencies in extracting desired information.
Innovation Solution
The method involves splitting an input feature map into sub-feature maps based on a dilation rate, performing convolution operations between these sub-feature maps and trained weights to generate intermediate feature maps, and then merging them to produce a dilated output feature map, which reduces the number of operations needed without increasing the number of parameters, thereby enhancing computing speed and power efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by moving object
If typical low-power neural network devices are used to process complex input data, then power consumption is reduced, but processing efficiency and real-time data extraction capability deteriorate
Solution Approach 1:
The input feature map is divided into multiple sub-feature maps based on dilation rate, allowing parallel processing of different regions. This segmentation enables the system to process complex data more efficiently while maintaining low power consumption by distributing computational load across multiple smaller processing units rather than one large unit.
2Measurement precision
If complex convolution operations are performed on large input data, then accurate information extraction is achieved, but computational requirements and processing time increase
Solution Approach 1:
By dividing the large input feature map into multiple sub-feature maps and processing them in parallel through multiple convolution operations, the system maintains accurate information extraction while significantly reducing overall processing time compared to sequential processing of the entire map.
Solution Approach 2:
The patent introduces dilation rate as an additional dimensional parameter to control the spacing of sampling points in the convolution operation. This allows the system to expand the receptive field and improve information extraction accuracy without increasing the number of parameters or computational complexity in the traditional sense.
3Adaptability or versatility
If dilation rate is increased to expand receptive field, then feature extraction capability is improved, but computational complexity increases
Solution Approach 1:
The patent segments the dilated convolution operation into multiple standard convolution operations applied to separate sub-feature maps. Each sub-feature map is processed with a smaller effective dilation rate, reducing the computational complexity of individual operations while collectively achieving the same expanded receptive field as a single high-dilation operation.
Data Source
AI summary
A neural network apparatus includes one or more processors configured to acquire an input feature map and trained weights, generate a plurality of sub-feature maps by splitting the input feature map based on a dilation rate, generate a plurality of intermediate feature maps by performing a convolution operation between the plurality of sub-feature maps and the trained weights, and generate a dilated output feature map by merging the plurality of intermediate feature maps based on the dilation rate.