AI Model Compression Control for Accuracy and Resource Trade-Offs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing artificial intelligence models, particularly those with a large number of neurons, consume significant resources and take a long time to operate due to their size, despite providing high accuracy in output values.
Innovation Solution
An electronic apparatus generates compressed artificial intelligence models by reducing the size of neural networks using methods like singular value decomposition (SVD) and reconstituting learning data sets to maintain performance, thereby optimizing resource usage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a vast amount of neurons are used in the deep neural network, then the accuracy of output values becomes high, but a lot of time is spent on operations and resources are consumed
Solution Approach 1:
The patent segments the deep neural network into multiple compressed neural networks with different compression rates. Each compressed network handles different operational requirements, allowing the system to select an appropriate level of compression based on the specific task, thus balancing accuracy and operation speed without sacrificing overall performance
Solution Approach 2:
The patent changes the parameter of neural network size by applying different compression rates to generate multiple versions of the network. This allows the system to operate with optimized parameters for specific tasks, reducing unnecessary computational overhead while maintaining required accuracy levels
2Measurement precision
If a vast amount of neurons are used in the deep neural network, then the accuracy of output values becomes high, but significant resources are consumed
Solution Approach 1:
The patent divides the resource-intensive deep neural network into multiple compressed versions, enabling the system to allocate computational resources more efficiently. Less complex compressed networks can handle simpler tasks with lower resource consumption, while more complex tasks can utilize the full-capacity original network only when necessary
Solution Approach 2:
The patent applies partial compression to the neural network, retaining only the essential components needed for specific tasks. This partial action approach reduces resource consumption by eliminating redundant neurons and connections while maintaining sufficient accuracy for practical applications
3Productivity
If the size of neural networks is reduced through compression, then resource usage is optimized, but the accuracy of output values may deteriorate
Solution Approach 1:
The patent creates a segmented hierarchy of compressed neural networks with varying compression rates. This segmentation allows the system to select the appropriate level of compression based on task requirements, ensuring that accuracy is maintained at levels sufficient for practical applications while optimizing resource usage
Solution Approach 2:
The patent employs feedback mechanisms to evaluate the performance of compressed neural networks and adjust compression parameters accordingly. By monitoring output accuracy and resource consumption, the system can iteratively optimize the compression rate to achieve the best balance between maintaining accuracy and reducing resource usage
Data Source
Figure 1
Figure 2
Figure 3A
AI summary
An electronic device control method is disclosed. The method for controlling an electronic device comprises the steps of: selecting a general-purpose artificial intelligence model; generating a reduced-artificial-intelligence model on the basis of the selected general-purpose artificial intelligence model; and generating an exclusive-use artificial intelligence model on the basis of the reduced-artificial-intelligence model, wherein the step of generating the reduced-artificial-intelligence model comprises a step of acquiring the rank of a singular value decomposition (SVD) algorithm on the basis of a reduction rate, reducing and training the selected general-purpose artificial intelligence model on the basis of the acquired rank so as to be converted into the reduced-artificial-intelligence model, determining the performance of the reduced-artificial-intelligence model on the basis of a preset first threshold, and generating the exclusive-use artificial intelligence model if the performance of the reduced artificial intelligence model is less than the preset first threshold.