CNN Feature Map Re-ordering for Data Traffic Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Convolutional neural networks (CNNs) face challenges in reducing memory access and internal network bandwidth consumption during image processing, leading to excessive data traffic and power consumption, particularly due to the high frequency of read/write operations for intermediate results and weights in feature classification layers.
Innovation Solution
A method is introduced that involves receiving multi-channel feature map images in a second layer of a CNN, analyzing their dynamic range, re-ordering them based on this range, and performing inter-channel prediction and encoding/decoding operations in a pipelined process with subsequent convolution operations, thereby reducing data traffic and power consumption.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by moving object
If multi-channel feature map images are processed in conventional CNN layers without re-ordering, then the processing follows standard computational flow, but data traffic and power consumption increase due to frequent memory read/write operations
Solution Approach 1:
The patent applies preliminary action by re-ordering multi-channel feature map images based on dynamic range analysis before they enter subsequent CNN layers. This re-ordering operation is performed in advance to optimize data distribution and minimize memory access operations during the actual convolutional processing, thereby reducing power consumption without affecting the fundamental CNN computation flow
2Loss of energy
If multi-channel feature map images are re-ordered based on dynamic range, then data traffic and power consumption are reduced, but additional processing steps are required
Solution Approach 1:
The patent applies parameter changes by analyzing and re-ordering feature map channels based on their dynamic range parameters. By changing the arrangement parameter of channels according to their dynamic characteristics, the system optimizes memory access patterns and reduces redundant data transmission, achieving lower power consumption while maintaining processing efficiency through parallel operations
3Productivity
If conventional CNN processing is used without dynamic range-based re-ordering, then the processing pipeline is simpler, but memory access frequency and data traffic are excessive
Solution Approach 1:
The patent applies preliminary action by re-ordering multi-channel feature map images based on dynamic range analysis before they enter subsequent CNN layers. This re-ordering operation is performed in advance to optimize data distribution and minimize memory access operations during the actual convolutional processing, thereby reducing power consumption without affecting the fundamental CNN computation flow
Solution Approach 2:
The patent applies parameter changes by analyzing and re-ordering feature map channels based on their dynamic range parameters. By changing the arrangement parameter of channels according to their dynamic characteristics, the system optimizes memory access patterns and reduces redundant data transmission, achieving lower power consumption while maintaining processing efficiency through parallel operations
Data Source
Figure 1
Figure 2
Figure 3~4B
AI summary
A convolutional neural network-based image processing method is provided. The method includes: receiving, in a second layer, multi-channel feature map images generated by applying a convolution operation to an input image of a convolutional neural network having a pluraltiy of layers with a plurality of filter kernels of a first layer; analyzing a dynamic range of the multi-channel feature map images; re-ordering the multi-channel feature map images, based on the dynamic range; and processing the re-ordered multi-channel feature map images in the second layer.