Global Pooling Method for Neural Networks Reducing Calculation Delay
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional Convolutional Neural Networks (CNNs) face issues with overfitting due to large parameter numbers in fully connected layers, leading to increased processing time and calculation delays, especially when global average pooling is applied to entire images at once, which wastes storage capacity and causes long calculation delays.
Innovation Solution
A global pooling method for neural networks that processes point data sequentially, performing a preset pooling operation on each piece of data as it is received, allowing for parallel storage and calculation, reducing the need for centralized pooling and thus minimizing calculation delay and storage requirements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of time
If global average pooling is applied to entire images at once, then pooling operation can be completed, but storage capacity is wasted and calculation delay increases
Solution Approach 1:
The patent divides the global pooling operation into multiple sequential steps, processing one pixel at a time through multiple network layers before moving to the next pixel. This segmentation allows the system to complete pooling operations incrementally rather than requiring all pixels to be stored and processed simultaneously, thereby reducing storage capacity requirements while maintaining calculation accuracy.
2Productivity
If fully connected layer is used in CNN, then network structure is complete, but parameter number increases leading to overfitting and increased processing time
Solution Approach 1:
The patent extracts and removes the fully connected layer from the conventional CNN structure, replacing it with a global pooling layer that processes data through multiple convolutional layers sequentially. This extraction eliminates the excessive parameters associated with fully connected layers, reducing the risk of overfitting and decreasing processing time while maintaining the network's ability to learn effective features.
Data Source
AI summary
Disclosed are a global pooling method for a neural network and a many-core system. The global pooling method for a neural network includes: receiving point data of to-be-processed data sequentially input by a previous network layer; and performing a preset pooling operation on the received point data after each piece of point data is received until the pooling operations of all the point data of the to-be-processed data are completed.