Object-Size-Aware Neural Network Pruning for Lightweight Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional global pruning techniques for neural network models face high calculation costs due to the increasing scale and computational complexity of deep neural networks, limiting their usability in resource-constrained environments, and there is a need for a method to compress these models with minimal performance loss.
Innovation Solution
A method that calculates pruning ratios based on object size information within a data set to optimize the compression of neural network models, using pruning ratios specific to each layer, thereby reducing calculation costs and maintaining performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional global pruning techniques use spectral norm to determine pruning ratio, then the pruning can be applied to the entire neural network, but the calculation costs increase significantly when the size of the deep neural network is large
Solution Approach 1:
The patent divides the neural network into multiple layers and applies different pruning ratios to each layer based on object size categories. Instead of using a single global spectral norm calculation, the method segments the pruning process by layer, allowing selective compression that reduces overall calculation costs while maintaining performance for different object scales
Solution Approach 2:
The patent implements local quality by assigning different pruning strategies to different layers based on their specific functionality and the size of objects they process. Each layer receives a customized pruning ratio tailored to its role in the network, optimizing the balance between compression and performance preservation for local regions of the model
2Measurement precision
If the scale and computational complexity of deep neural network models are increased for high performance, then the recognition accuracy is improved, but the usability and accessibility in situations where hardware resources are limited are constrained
Solution Approach 1:
The patent implements dynamic adaptability by enabling the neural network to adjust its effective complexity based on input characteristics. Through object size-aware pruning, the model dynamically allocates computational resources to the most relevant features for the given input, achieving high accuracy when needed while reducing resource consumption for simpler cases
Solution Approach 2:
The patent changes the parameters of the neural network by applying different pruning ratios to different layers based on object size categories. This parameter adjustment allows the model to optimize its structure for specific task requirements, reducing the effective number of parameters and computations while maintaining recognition accuracy for the target application scenario
3Reliability
If pruning ratio is set globally for the entire neural network, then the risk of losing important information is reduced, but the calculation costs increase when the size of the deep neural network is large
Solution Approach 1:
The patent segments the neural network into layers processing different object size categories and applies targeted pruning to each segment. This segmentation allows the preservation of important information in layers handling critical object sizes while applying more aggressive pruning to layers processing less critical sizes, reducing overall calculation costs without compromising reliability
Data Source
AI summary
A method of compressing a neural network model for object recognition according to an embodiment of the present application includes: receiving an original model for object recognition trained based on a first data set; receiving a second data set of an analysis target; calculating object size information on sizes of objects included in an image of the second data set; performing pruning on at least one layer included in the original model based on the calculated object size information; and generating a compressed neural network model from the original model based on the results of performing the pruning.