Edge-Based Learning for Real-Time AI Classification on IoT Devices
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Sophisticated AI/ML models require significant processing resources and time, making them unsuitable for real-time applications on edge devices like IoT devices, which lack the necessary computational power and memory.
Innovation Solution
A heterogeneous AI/ML architecture is employed, where complex AI/ML techniques are used in conjunction with lightweight techniques, with edge devices leveraging classifications and labels provided by cloud-based systems to perform edge-based learning, allowing for accurate data classification with reduced processing power and time requirements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If sophisticated AI/ML models are used for accurate data classification, then classification accuracy is improved, but processing resource requirements and time increase significantly
Solution Approach 1:
The patent segments the AI/ML processing into two distinct parts: a complex teacher model for training that achieves high accuracy, and a simplified student model for deployment that requires minimal resources. The teacher model processes data to generate training examples, which are then used to train the lightweight student model, effectively dividing the computational burden between training and inference phases.
Solution Approach 2:
The patent creates a simplified copy (student model) of the complex teacher model. The student model is trained using distillation techniques to replicate the teacher model's classification behavior without requiring the same computational resources. This copying approach allows the system to maintain high accuracy while reducing device complexity for real-time inference.
2Measurement precision
If sophisticated AI/ML models are used for accurate data classification, then classification accuracy is improved, but processing time increases making real-time applications infeasible
Solution Approach 1:
The patent separates the time-consuming training phase (performed by the teacher model offline) from the rapid inference phase (performed by the student model in real-time). This segmentation allows accurate classification to be achieved without real-time processing delays, as the student model has already learned the classification patterns during offline training.
Solution Approach 2:
The patent performs preliminary action by pre-training the student model using examples generated by the teacher model before deployment. This preliminary training phase prepares the lightweight model in advance, enabling it to perform accurate real-time classification without requiring complex computations during actual inference, thus reducing processing time for real-time applications.
3Measurement precision
If complex AI/ML techniques are deployed on edge devices, then model accuracy can be maintained, but computational constraints of edge devices are exceeded
Solution Approach 1:
The patent introduces a cloud-based teacher model as an intermediary that generates training data for edge devices. The teacher model performs complex computations in the cloud where computational power is abundant, then distills knowledge into simplified student models that can run on resource-constrained edge devices. This intermediary approach allows edge devices to maintain accuracy without requiring high computational power locally.
Solution Approach 2:
The patent creates a lightweight copy (student model) of the complex teacher model that is optimized for deployment on edge devices. The copying process involves distilling the essential classification knowledge into a smaller, more efficient model structure that maintains accuracy while adapting to the computational constraints of edge devices with limited power and processing capabilities.
Data Source
AI summary
Techniques described herein provide for the use of a heterogeneous artificial intelligence/machine learning (“AI/ML”) architecture, in which relatively complex AI/ML, techniques may be used in conjunction with more lightweight AI/ML techniques in order to leverage the accuracy of relatively complex AI/ML techniques with the reduced processing power and/or time requirements of more lightweight AI/ML techniques. A teacher model system may utilize processing resource and/or time-intensive AI/ML techniques and/or models in order to determine classifications associated with source data, and may provide such classifications to a student model system that may utilize the classifications in accordance with less processing resource and/or time-intensive AI/ML techniques in order to accurately classify sensor data in real time or near-real time.