Supernet Model Space for Edge Device Deployment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Edge computing devices, such as laptops and IoT devices, have limited memory and processing power, making it difficult to train deep learning models locally without substantial computational resources.
Innovation Solution
A method to deploy deep learning models on edge devices by creating a supernet model space and training multiple models within it, then selecting and deploying the appropriate model based on the device's characteristics, such as memory and processing power, without the need for retraining or fine-tuning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If deep learning models are trained on edge devices, then model performance can be optimized for local tasks, but the limited memory and processing power of edge devices make training infeasible
Solution Approach 1:
A server acts as an intermediary between the training data and edge devices. The server performs the computationally intensive training of multiple specialized models using a supernet architecture, then distributes these pre-trained models to edge devices. This mediator approach allows edge devices to benefit from optimized models without requiring local training capabilities.
Solution Approach 2:
The system performs preliminary training actions on the server before deployment to edge devices. Multiple specialized models are trained in advance using the supernet framework, and the results are cached and made available for immediate deployment. This eliminates the need for time-consuming local training and allows edge devices to use pre-optimized models.
2Adaptability or versatility
If multiple specialized models are trained for different edge devices, then each device gets an optimized model, but the training process becomes complex and resource-intensive
Solution Approach 1:
The supernet architecture provides a universal model structure that can serve multiple different edge devices with varying requirements. A single supernet is trained to encompass multiple specialized models, allowing the system to generate device-specific models from a unified training process. This multi-functional approach simplifies the overall system while maintaining device-specific optimization.
Solution Approach 2:
The training processes for multiple specialized models are merged into a single supernet training operation. Instead of training separate models independently, the system combines them into one unified training process that efficiently produces all specialized models simultaneously, reducing overall complexity and resource requirements.
3Productivity
If pre-trained models are deployed to edge devices, then deployment time is reduced, but the models may not be optimized for specific device characteristics
Solution Approach 1:
The system generates locally optimized models for each edge device based on its specific characteristics such as memory capacity, processing power, and hardware architecture. Each device receives a customized model from the supernet that is tailored to its local conditions, ensuring optimal performance while maintaining fast deployment through the pre-training approach.
Data Source
AI summary
Computer-implemented methods for deploying models at edge devices without retraining. Aspects include receiving, from a user an identification of a computing task to be performed by an edge device and obtaining, a data set corresponding to the computing task. Aspects also include determining a supernet model space based at least in part on the computing task and creating a plurality of trained models for the computing task by training a plurality of deep learning models within the supernet model space with the data set. Aspects further include deploying one of the plurality of trained models to the edge device, wherein the one of the plurality of trained models is determined based at least in part on one or more characteristics of the edge device.