Partitioned Machine Learning Model for Edge Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models, particularly deep learning models, face challenges in being implemented on resource-constrained platforms like drones, robots, and smartphones due to high resource overhead, leading to the need for offloading training and execution to remote cloud platforms, which incurs significant communication overhead and compromises data security.
Innovation Solution
A partitioned machine learning architecture that divides a global machine learning model into local models based on resource constraints, allowing these models to be trained and executed on the device, with data preprocessing to reduce dimensionality and enable efficient processing, while minimizing communication costs and preserving data privacy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning models are implemented on resource-constrained platforms, then inference accuracy is improved, but resource overhead increases
Solution Approach 1:
The global machine learning model is partitioned into multiple local machine learning models that can be executed on resource-constrained platforms. Each local model handles a subset of the computational tasks, enabling deep learning inference accuracy while reducing the resource overhead by distributing the model across multiple smaller components that fit within device memory and processing capabilities.
2Quantity of substance
If training and execution are offloaded to remote cloud platforms, then resource overhead is reduced, but communication overhead increases
Solution Approach 1:
The model is segmented into local and global components, where local models are deployed on edge devices for on-device inference, eliminating the need to offload every inference to the cloud. This reduces communication overhead while maintaining the ability to use cloud resources for model updates and retraining when needed.
Solution Approach 2:
Local machine learning models enable devices to perform inference autonomously without requiring constant cloud connectivity. The device can independently process data locally, reducing communication overhead and energy loss associated with continuous cloud interactions, while still benefiting from periodic model updates.
3Quantity of substance
If data is transmitted to cloud platforms for processing, then resource overhead is reduced, but data security is compromised
Solution Approach 1:
By partitioning the model and keeping local copies on edge devices, sensitive data can remain on-device rather than being transmitted to the cloud. Only model parameters and aggregated insights need to be communicated, preserving data security while still enabling access to powerful processing capabilities when needed.
Data Source
AI summary
A system may include a processor and a memory. The memory may include program code that provides operations when executed by the processor. The operations may include: partitioning, based at least on a resource constraint of a platform, a global machine learning model into a plurality of local machine learning models; transforming training data to at least conform to the resource constraint of the platform; and training the global machine learning model by at least processing, at the platform, the transformed training data with a first of the plurality of local machine learning models.