Neural Network Parameter Sharing for Hardware Adaptability
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural network training methods require significant processing resources and often result in sub-optimal performance due to the lack of consideration for specific hardware environments during design, leading to inefficiencies in inference processes.
Innovation Solution
A method is introduced that identifies and optimizes a group of neural networks, including a main neural network and sub-neural networks with shared parameters, by iteratively adjusting and scoring their performance across various hardware environments, allowing for selection of the most suitable networks for processing data in multiple hardware environments.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a single neural network is trained using centralized approaches with significant processing resources, then the neural network can be deployed for inference, but the performance is sub-optimal due to lack of consideration for specific hardware environments
Solution Approach 1:
The patent segments the training process by creating multiple specialized neural networks (first, second, and third neural networks) optimized for different hardware environments (first, second, and third hardware environments respectively). Each network is trained separately with hardware-specific optimizations, allowing the system to select the appropriate network for the target hardware, thereby resolving the contradiction between achieving high performance and maintaining hardware adaptability.
2Productivity
If neural networks are optimized for specific hardware environments, then inference efficiency improves, but the complexity of providing and selecting appropriate networks increases
Solution Approach 1:
The patent applies preliminary action by pre-training multiple neural networks for different hardware environments before deployment. The system prepares all necessary neural networks in advance, each optimized for specific hardware characteristics, so that at inference time, the system simply needs to select the pre-prepared network matching the target hardware, thereby improving inference efficiency without adding runtime complexity.
3Reliability
If multiple neural networks are trained for different hardware environments, then hardware-specific optimization is achieved, but the processing resources required for training increase
Solution Approach 1:
The patent applies local quality by training each neural network with hardware-specific characteristics and optimizations tailored to its target environment. Instead of using a generic training approach for all networks, each network receives localized training considerations specific to its hardware platform, achieving better hardware optimization while allowing distributed or selective training to manage resource consumption.
Data Source
AI summary
A computer-implemented method of providing a group of neural networks for processing data includes: identifying a group of neural networks including a main neural network and one or more sub-neural networks, each neural network comprising a plurality of parameters and wherein one or more of the parameters of each sub-neural network are shared by the sub-neural network and the main neural network; inputting training data into each neural network, and adjusting the parameters of each neural network; computing a performance score for each neural network using the adjusted parameters; generating a combined score for the group of neural networks by combining the performance score, with a value of a loss function computed for each neural network using the adjusted parameters; repeating the identifying and the inputting and the adjusting and the computing and the generating; and selecting a group of neural networks for processing data in the plurality of hardware environments based on the value of the combined score for each group of neural networks.