Mobile AI Model Pruning for Low-Power Device Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models struggle to automatically configure and deploy on low-power devices while ensuring accuracy and resource performance, particularly in fields with limited training data and hardware constraints.
Innovation Solution
A method involving pruning layers from a machine learning model based on device specifications, training a student model with teacher models, and adjusting layers to meet performance requirements, using both labeled and unlabeled data for retraining.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by moving object
If a machine learning model is configured to function within hardware specifications of low-power devices, then resource overhead is reduced, but performance (accuracy) suffers
Solution Approach 1:
The model is divided into multiple layers, and only a subset of layers is selected and retained for deployment on low-power devices. This segmentation allows the system to keep only the most essential computational components, reducing resource overhead while maintaining sufficient accuracy for the target application.
Solution Approach 2:
Different layers of the model are evaluated individually to determine their contribution to overall performance. Layers are selectively retained or removed based on their local importance to the task, allowing optimization tailored to specific hardware constraints while preserving critical functionality.
2Device complexity
If a lightweight machine learning model is used, then hardware resource requirements are reduced, but accuracy typically suffers
Solution Approach 1:
Instead of using the complete original model, only a partial subset of layers is deployed on the low-power device. This partial action approach reduces model complexity and resource requirements while maintaining adequate accuracy by retaining the most critical layers needed for the specific task.
Solution Approach 2:
The system changes the parameter of model size by selectively removing layers, transforming a complex full-model into a simplified lightweight version. This parameter change allows the model to adapt to hardware constraints while preserving essential predictive capabilities through intelligent layer selection.
3Quantity of substance
If layers are deleted from a machine learning model, then resource overhead is reduced, but performance may deteriorate
Solution Approach 1:
The system performs preliminary evaluation of each layer's contribution to model performance before making deletion decisions. By assessing layer importance in advance, the system can confidently remove layers that have minimal impact on accuracy, thereby reducing the number of layers and resource overhead without significantly deteriorating model performance.
Solution Approach 2:
The system uses performance feedback from evaluating pruned model variants to guide further layer deletion decisions. By monitoring how performance changes after each pruning step, the system can identify the optimal point where resource overhead is minimized while performance remains acceptable for the target application.
Data Source
AI summary
A machine learning model can be optimized for deployment on a device based on hardware specifications of the device. An existing model is acquired and pruned to reduce hardware resource consumption of the model. The pruned model is then trained based on training data. The pruned model is also trained based on a collection of “teacher” models. Performance of the trained model is then evaluated and compared to performance requirements, which can be based on the hardware specifications of a device.