Pruned Neural Network Convolution for Resource-Limited Processors
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural network models consume significant computing resources and time for convolution operations, making them inefficient in resource-limited environments such as sensors and embedded systems.
Innovation Solution
Implement a processor-based method that utilizes pruning information to perform an accumulation type convolution operation, focusing on valid weight element vectors and reducing unnecessary computations through pointwise pruning, thereby optimizing resource usage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional convolution operations are performed using complete weight kernels, then comprehensive feature extraction is achieved, but computing resource consumption increases significantly
Solution Approach 1:
The patent extracts and removes redundant or insignificant weight elements from the complete weight kernel through pruning operations. This creates a sparse weight kernel that retains only the most important connections, thereby reducing computing resource consumption while preserving the essential feature extraction capability.
Solution Approach 2:
The patent applies different treatment to different regions of the weight kernel by identifying and pruning specific weight elements based on their importance. This creates a non-uniform weight distribution where important connections are preserved and less important ones are removed, optimizing the balance between accuracy and efficiency.
2Measurement precision
If conventional convolution operations process all weight elements, then complete feature mapping is achieved, but processing time increases
Solution Approach 1:
The patent extracts only the necessary weight elements that contribute meaningfully to the output. By pruning redundant weight elements beforehand, the convolution operation processes fewer elements, directly reducing processing time while maintaining output accuracy through selective retention of important weights.
Solution Approach 2:
The patent performs weight pruning as a preliminary action before the convolution operation. This pre-processing step identifies and removes insignificant weight elements in advance, so that the subsequent convolution operation only needs to process the essential remaining elements, thereby reducing processing time without compromising accuracy.
3Productivity
If pruning is applied to reduce weight kernel size, then computing efficiency improves, but model accuracy may deteriorate
Solution Approach 1:
The patent applies pruning selectively to individual weight elements based on their specific importance rather than uniformly reducing the entire weight kernel. This local quality approach ensures that important connections are preserved while less important ones are removed, maintaining accuracy while improving efficiency.
Solution Approach 2:
The patent changes the parameters of the weight kernel by adjusting which weight elements are retained and which are pruned. This parameter change is guided by importance metrics, allowing the model to adapt its weight configuration to achieve better efficiency-accuracy tradeoff without fundamentally altering the network architecture.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A processor-implemented neural network method includes: obtaining a first weight kernel of a weight model and pruning information of the first weight kernel; determining, based on the pruning information, a processing range of an input feature map for each weight element vector of the first weight kernel; performing a convolution operation between the input feature map and the first weight kernel based on the determined processing range; and generating an output feature map of a neural network layer based on an operation result of the convolution operation.