Class-Specific Surrogate Neural Networks for Resource-Limited Edge Devices
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural networks are resource-intensive and complex, making them unsuitable for execution on resource-constrained devices like edge devices without additional hardware, which is often not feasible due to energy consumption and resource constraints.
Innovation Solution
A method to create surrogate neural networks by splitting an initial neural network into subsets, retraining them to focus on specific classes, and optimizing them for resource-constrained devices, allowing for accurate predictions while reducing computational and energy demands.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a complex neural network is used to handle more classes and features, then prediction accuracy and functionality are improved, but resource consumption and device complexity 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 accurate predictions for specific classes while reducing the computational burden and energy consumption compared to running a single complex network for all classes.
Solution Approach 2:
Instead of training and deploying a complete complex neural network for all classes, the patent applies partial action by creating surrogate networks that handle only specific subsets of classes. This partial approach reduces resource consumption while maintaining sufficient prediction accuracy for the targeted classes.
2Adaptability or versatility
If a complex neural network is deployed on resource-constrained devices, then functionality is improved, but device resource requirements exceed available capacity
Solution Approach 1:
The complex neural network is segmented into multiple smaller surrogate networks, each with reduced complexity suitable for deployment on resource-constrained edge devices. This allows the system to maintain versatility across different classes while ensuring each individual network can run on limited hardware resources.
Solution Approach 2:
The patent creates copies of the neural network architecture and retrains them with specialized data subsets. These copied and specialized surrogate networks have reduced complexity requirements compared to the original complex network, enabling deployment on edge devices with limited resources.
3Power
If additional hardware like GPU is added to system, then neural network execution capability is improved, but energy consumption and resource constraints are violated
Solution Approach 1:
Instead of investing in expensive additional hardware like GPUs, the patent uses software-based surrogate neural networks that are computationally lighter. These surrogate networks provide sufficient execution capability for edge devices without requiring expensive hardware upgrades, thus avoiding increased energy consumption.
4Device complexity
If neural network size is reduced for edge deployment, then resource requirements are reduced, but prediction accuracy deteriorates
Solution Approach 1:
Each surrogate neural network is specialized for a specific subset of classes, achieving high prediction accuracy for its designated classes. This local quality approach ensures that while individual networks are smaller, they maintain or improve accuracy for their specific functionality compared to a generic reduced network.
Solution Approach 2:
The surrogate neural networks are pre-trained using ground truth predictions from the complex neural network as training data. This preliminary training action ensures that the smaller networks achieve high prediction accuracy for their specific classes without requiring the full complexity of the original network.
Data Source
AI summary
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.