Class-Specific Surrogate Neural Networks for Resource-Constrained Edge Devices
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural networks are resource-intensive and unsuitable for execution on resource-constrained devices like edge devices due to energy consumption and hardware limitations, particularly in embedded systems and edge computing scenarios.
Innovation Solution
A method to split a complex neural network into surrogate neural networks, each trained for a subset of classes, using the initial neural network's predictions as ground truth during retraining, and optimizing them for resource-efficient execution on edge devices.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a complex neural network is used to handle more classes and features, then the classification capability and accuracy are improved, but the energy consumption and computational resources increase
Solution Approach 1:
The patent divides a complex neural network into multiple specialized surrogate neural networks, each handling a subset of classes. This segmentation allows the system to achieve the same overall classification capability while reducing the computational burden and energy consumption of each individual network execution.
Solution Approach 2:
Instead of always deploying the full complex neural network, the system uses surrogate networks that handle only the necessary subset of classes for each prediction task. This partial action approach reduces computational resources and energy consumption while maintaining adequate classification capability for the specific task at hand.
2Productivity
If additional hardware like GPU is added to the system, then the neural network execution performance is improved, but the device complexity and resource requirements increase
Solution Approach 1:
The patent segments the neural network processing into multiple smaller surrogate networks that can execute on standard CPU hardware without requiring additional GPU resources. This segmentation enables high execution performance to be achieved through efficient software architecture rather than additional hardware.
Solution Approach 2:
The system creates surrogate copies of the original neural network, each trained on specific subsets of data. These surrogate copies can be deployed and executed on resource-constrained devices without requiring the original complex hardware infrastructure, thus improving productivity while reducing device complexity.
3Measurement precision
If the neural network architecture is made bigger to detect more features, then the measurement precision and classification accuracy are improved, but the computational overhead increases
Solution Approach 1:
The patent segments the feature detection and classification tasks across multiple surrogate neural networks, each optimized for specific subsets of classes. This segmentation maintains high measurement precision for each subset while reducing the computational overhead compared to a single large network processing all features simultaneously.
Solution Approach 2:
The system changes the parameters of the neural networks by training surrogate models with different architectures and hyperparameters optimized for their specific subsets of classes. This allows each surrogate network to achieve high classification accuracy for its domain while maintaining lower computational overhead than a universal large-scale network.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The invention provides a computer-implemented method for providing one or more surrogate neural networks for execution on resource-constrained device, such as an edge device, the method comprising retrieving a trained initial neural network trained to make predictions for a set of classes of input data, selecting a subset of classes among the set of classes, the subset comprising one or more classes, creating a copy of the initial neural network, obtaining a surrogate neural network, the obtaining comprising retraining the copy of the initial neural network to make predictions for the subset of classes, wherein, for the retraining, predictions of the trained initial neural network are used as ground truth.