Neural Network Resolution Selection for Mobile Inference Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The challenge of performing training and inference of large-scale deep learning models on lightweight devices with limited computing resources, such as mobile devices, due to the high computational demands of modern neural networks.
Innovation Solution
Selecting neural networks based on minimal mutual information and configuring them to perform inference operations efficiently by considering limiting conditions such as hardware specifications, inference time, and accuracy requirements, allowing for parallel processing of neural networks with varying resolutions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If large-scale deep learning models are used to improve inference accuracy, then accuracy is improved, but computing resource consumption increases
Solution Approach 1:
The patent divides the large-scale deep learning model into multiple smaller neural networks with different input image resolutions. Each neural network processes images at its specific resolution, and their outputs are combined to achieve the final inference result. This segmentation allows the system to maintain high accuracy while reducing the computational burden on lightweight devices.
2Measurement precision
If multiple neural networks are used to improve inference accuracy, then accuracy is improved, but device complexity increases
Solution Approach 1:
The patent implements dynamic selection of neural networks based on the input image resolution. The system determines which neural networks to activate based on the actual input characteristics, rather than always running all networks. This dynamic approach reduces the effective complexity at inference time while maintaining the accuracy benefits of multiple networks.
3Adaptability or versatility
If neural networks are optimized for different resolutions to improve adaptability, then adaptability is improved, but mutual information increases
Solution Approach 1:
The patent changes the resolution parameter of neural networks to create a diverse set of models. By training neural networks at different resolutions, the system achieves better adaptability to various input sizes. The mutual information calculation is used as a selection criterion to choose networks that provide complementary information, balancing adaptability with information efficiency.
Data Source
AI summary
A method and apparatus with neural network control are provided. In one general aspect, a method is performed by at least one processor of an apparatus, the method includes selecting neural networks from among available neural networks with respective input image resolutions, including plural resolutions, wherein the selected neural networks are selected based on a sum of amounts of mutual information between the selected neural networks being minimal and based on a limiting condition of a use of the available neural networks, and configuring the available neural networks to perform an inference on input data, wherein the configuring is based on the combination of the selected neural networks.