Hierarchical Transfer Learning for Edge Model Adaptation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Training deep learning models on a single device or a small number of devices is computationally expensive and impractical due to the need for large training datasets and specialized processors, making it difficult to develop models that effectively perform tasks on edge devices without extensive resources.
Innovation Solution
A method that involves training a central deep learning model using a large dataset and then customizing it for edge devices using transfer learning, deploying the customized model to edge devices through a hierarchy of models including a central common model, location-based models, and device-based models, leveraging cloud computing for resource efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a deep learning model is trained on a single device or a small number of devices, then the model can be customized for specific device characteristics, but the computational cost and resource requirements become excessively high and impractical
Solution Approach 1:
The patent segments the model training process into multiple hierarchical levels: a central common model trained on aggregated data from all devices, location-based models trained on data from devices within specific geographic areas, and device-based models trained on data from individual devices. This segmentation allows the system to distribute computational workload across different levels, reducing the computational burden on any single device while maintaining customized models for specific device characteristics.
Solution Approach 2:
The patent introduces an intermediary cloud-based training system that mediates between individual edge devices and the final customized models. The cloud platform aggregates data from multiple devices, trains the hierarchical models, and deploys customized models to edge devices. This intermediary approach eliminates the need for each edge device to perform computationally intensive training operations locally, significantly reducing individual device resource requirements.
2Reliability
If a deep learning model is trained using large training datasets, then the model performance improves, but the requirement for specialized processors and extensive resources makes deployment on edge devices difficult
Solution Approach 1:
The patent divides the training dataset into multiple subsets corresponding to different hierarchical levels: data from all devices for the central common model, data from devices within location proximity for location-based models, and data from individual devices for device-based models. This segmentation enables the system to train models with sufficient accuracy using appropriately scaled datasets for each level, avoiding the need for edge devices to process entire large datasets while maintaining high model performance.
Solution Approach 2:
The patent transitions from a single-level training approach to a multi-dimensional hierarchical training architecture. Instead of training one large model requiring extensive resources, the system creates multiple models at different levels of abstraction (central, location-based, device-based). This dimensional change in the training architecture allows edge devices to deploy lightweight customized models while the cloud handles the computationally intensive aggregation and training of broader datasets.
Data Source
AI summary
A computer-implemented method for adapting a deep learning model to a local environment includes collecting training data; training a common deep learning model using the collected training data; customizing the deep learning model based on characteristics specific to one of a plurality of local devices utilizing transfer learning; and deploying the customized deep learning model to the one of the plurality of local devices. The customized deep learning model may be trained across a hierarchy comprising: a central common model trained on data from all devices in the plurality of devices; a location based model trained on a first set of data from devices in the plurality of devices within a location proximity; and a device based model trained on a data from one device in the plurality of devices.