REC Model Segmentation for Edge ML Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional machine learning models for classification and anomaly detection have large memory footprints, long training times, and high power consumption, making them unsuitable for implementation on resource-constrained edge devices in IoT deployments.
Innovation Solution
A lightweight and efficient ML model, referred to as the REC model, comprising a primary ML model and secondary ML models, where the primary model handles easy data instances with high confidence and forwards uncertain instances to secondary models for further classification, reducing memory consumption and latency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional monolithic classifiers are used for classification, then classification accuracy can be maintained, but memory footprint increases and training time becomes longer
Solution Approach 1:
The patent divides the monolithic classifier into multiple specialized classifiers (e.g., normal class classifier and anomaly class classifier). Each classifier is trained on specific data subsets and handles particular classification scenarios, reducing the memory requirements of individual classifiers while maintaining overall classification accuracy through ensemble decision-making.
2Reliability
If traditional monolithic classifiers are used for classification, then classification accuracy can be maintained, but training time increases and power consumption rises
Solution Approach 1:
The training process is segmented into multiple smaller training tasks, each focused on specific data subsets and class characteristics. This allows for more efficient training convergence and better utilization of available computational resources, reducing overall training time and power consumption while achieving comparable accuracy to monolithic classifiers.
3Adaptability or versatility
If traditional monolithic classifiers are used for classification, then comprehensive classification capability is achieved, but classification latency increases
Solution Approach 1:
The classification process is segmented into multiple specialized classifiers that can be evaluated in parallel or sequentially based on data characteristics. This segmentation enables faster decision-making for certain data types while maintaining comprehensive classification coverage, thereby reducing overall classification latency.
Solution Approach 2:
The system employs multiple classifiers beyond what a single monolithic classifier would provide, with each classifier performing partial classification tasks. This excessive action approach ensures that even if one classifier is uncertain, other classifiers can compensate, maintaining high classification capability while enabling faster throughputs via parallel evaluation.
Data Source
AI summary
Techniques for implementing an efficient machine learning (ML) model for classification are provided. In one set of embodiments, a computer system can receive a query data instance to be classified. The computer system can then generate a first classification result for the query data instance using a first (i.e., primary) ML model, where the first classification result includes a predicted class for the query data instance and a confidence level indicating a likelihood that the predicted class is correct, and compare the confidence level with a classification confidence threshold. If the confidence level is greater than or equal to the classification confidence threshold, the computer system can output the first classification result as a final classification result for the query data instance. However, if the confidence level is less than the classification confidence threshold, the computer system can forward the query data instance to one of a plurality of second (i.e., secondary) ML models for further classification.