CNN Filter Pruning with Adaptive Factors for Edge Deployment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Convolutional neural networks (CNNs) require high computing power and resources, making them unsuitable for edge devices with limited resources, and existing pruning methods are not universally effective across different applications.
Innovation Solution
A method for filter-based pruning of CNNs that adjusts filter factors during training, allowing simultaneous pruning and training, and optimizes the number of filters based on processing units to maximize resource utilization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by moving object
If filter-based pruning is applied to reduce computational requirements, then the CNN becomes suitable for edge devices with limited resources, but the number of filters is reduced which may affect processing capability
Solution Approach 1:
The patent introduces a filter factor parameter that dynamically controls the activation and deactivation of filters during CNN operation. By adjusting this parameter, the system can reduce computational requirements by deactivating less important filters while maintaining the structural integrity of the network, thus resolving the contradiction between reducing energy use and maintaining processing capability
Solution Approach 2:
The patent implements dynamic filter selection where the importance of each filter is evaluated in real-time based on the input data and current operational context. This allows the CNN to adaptively activate only the necessary filters for each inference task, reducing computational load while preserving processing capability when needed
2Productivity
If the number of filters is optimized to match SIMD processing units, then resource utilization is maximized, but the flexibility to handle different application requirements is reduced
Solution Approach 1:
The patent segments the filter set into multiple groups that can be independently activated or deactivated. This segmentation allows the system to optimize resource utilization by activating only the necessary number of filter groups to match SIMD processing unit capacity, while retaining the ability to activate additional groups if needed for different application requirements
Solution Approach 2:
The patent designs the filter structure to be universally applicable across different applications and hardware configurations. The same filter framework can be configured to work with different numbers of SIMD processing units by dynamically adjusting which filters are activated, thus achieving both optimized resource utilization and maintained flexibility
Data Source
AI summary
A system and method for filter based pruning of a convolutional neural network (CNN), is disclosed. The method includes initializing a CNN, the CNN including a plurality of filters, each filter associated with a weight and a filter factor; providing the CNN with a training input; adjusting a weight of a filter of the plurality of filters in response to processing the training input; adjusting a filter factor of the filter of the plurality of filters in response to processing the training input; pruning the CNN by removing the filter in response to detecting that a value of the filter factor is below a predefined threshold after training is complete; storing a trained pruned CNN based on the initialized CNN; and processing an input with the trained CNN.