Edge AI Classification via Microcontroller Thresholding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing AI systems require resource-intensive hardware and costly network bandwidth for visual recognition operations, making them inefficient and expensive for performing AI classification and detection tasks.
Innovation Solution
A method for edge classification using artificial intelligence that employs low-end microcontrollers to capture and classify data locally, using AI algorithms to predict and compare results against a threshold, thereby minimizing the need for high-level AI systems and reducing network bandwidth usage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If high-level AI systems are used for all classification tasks, then accuracy is improved, but network bandwidth consumption and operating costs increase
Solution Approach 1:
The patent segments AI processing into two levels: edge devices perform initial classification of simple tasks, while server assembly handles complex analysis. This segmentation allows accurate classification to be achieved for routine tasks at the edge, eliminating the need to transmit all data over the network, thus reducing network bandwidth consumption while maintaining accuracy for tasks that require it.
Solution Approach 2:
The patent introduces an intermediary threshold comparison mechanism between edge devices and server assembly. Edge devices classify data locally and only transmit results that fail to meet the accuracy threshold to the server assembly. This intermediary filtering approach ensures high accuracy for transmitted data while minimizing network bandwidth usage by filtering out successfully classified data.
2Measurement precision
If high-level AI systems are used for all tasks, then detection accuracy is improved, but upfront hardware costs increase
Solution Approach 1:
The patent segments the AI system into edge devices with basic AI capabilities and a server assembly with advanced AI capabilities. This segmentation allows organizations to deploy inexpensive edge devices for the majority of classification tasks, achieving sufficient accuracy for simple tasks, while reserving expensive high-level AI systems only for complex analysis, thus reducing upfront hardware costs while maintaining detection accuracy where needed.
Solution Approach 2:
The patent employs inexpensive edge devices that can be deployed widely for routine classification tasks. These low-cost devices handle the majority of AI workloads, eliminating the need to deploy expensive high-level AI systems everywhere. The threshold-based filtering ensures that only cases requiring expensive server resources are transmitted, making the overall system more cost-effective while maintaining accuracy.
3Loss of energy
If edge devices perform AI classification, then network bandwidth is reduced, but device complexity increases
Solution Approach 1:
The patent implements partial AI processing at the edge, where devices perform only the classification function with a threshold comparison, rather than implementing complete AI analysis. This partial action approach reduces edge device complexity by limiting their role to basic classification and threshold checking, while still achieving network bandwidth reduction by filtering out successfully classified data before transmission.
4Loss of energy
If simple AI tasks are handled by low-power systems, then operating costs are reduced, but system reliability may worsen
Solution Approach 1:
The patent implements a feedback mechanism where edge devices transmit their classification results and confidence levels to the server assembly. The server assembly provides feedback by processing cases that fall below the accuracy threshold and can update or correct edge device classifications. This feedback loop ensures system reliability is maintained by allowing the more powerful server to correct or validate edge device decisions, preventing reliability degradation despite using lower-power systems for simple tasks.
Data Source
AI summary
Disclosed herein is a method for an accurate edge classification using artificial intelligence. The system includes a network and a plurality of terminal devices communicably coupled via the network. The plurality of terminal devices are configured to capture data. The system further includes microcontrollers communicably coupled to the plurality of terminal devices. The microcontrollers are configured to send data from the plurality of terminal devices to the microcontrollers via the network. The microcontrollers are further configured to classify the received data and predict the classified data via the microcontrollers using Artificial Intelligence (AI) algorithms. The microcontrollers are further configured to compare the prediction against a threshold using the microcontrollers, wherein if the prediction is below the threshold, nothing is detected, and if the prediction is within the threshold, data is transmitted to higher-level/compute AI algorithms running on server assembly, else when the prediction is higher than the threshold data is determined as accurate.