Multi-Stage Neural Network Classifier for Edge Devices
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Large machine-learning models for event classification require significant computational resources and energy, making them unsuitable for deployment in resource-constrained settings like wearable devices.
Innovation Solution
A multi-stage machine learning classifier system that employs a first stage classifier to identify potential events of interest in sensor data segments, which are then transmitted to a second stage classifier for full classification, reducing computational and energy demands.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a large machine-learning model is used for event classification, then classification accuracy is improved, but computational resource requirements and energy consumption increase significantly
Solution Approach 1:
The patent divides the classification task into two stages: a first stage classifier that performs initial filtering and a second stage classifier that performs detailed classification. This segmentation allows the system to use a smaller, less energy-consuming model for the majority of cases while reserving the larger, more accurate model only for ambiguous cases, thus resolving the contradiction between accuracy and energy consumption.
Solution Approach 2:
The first stage classifier performs preliminary classification and filtering before the data reaches the second stage classifier. By pre-processing the data and eliminating clear cases early, the system avoids the high energy cost of running large models on all inputs, while still maintaining high overall accuracy through the two-stage approach.
2Measurement precision
If a large machine-learning model is deployed, then classification capability is improved, but device resource constraints are violated
Solution Approach 1:
The classification system is segmented into two distinct stages with different complexity levels. The first stage uses a simpler model suitable for resource-constrained devices, while the second stage handles complex classification only when necessary. This segmentation enables deployment on resource-constrained devices while maintaining high classification capability.
Solution Approach 2:
The system applies partial classification action at the first stage, handling only the necessary filtering and simple cases. The full classification capability is applied selectively at the second stage only when the first stage determines it is needed, thus avoiding the excessive resource consumption that would result from applying full classification capability to all inputs.
3Measurement precision
If all data segments are processed by a comprehensive classifier, then complete classification is achieved, but computational processing power requirements increase
Solution Approach 1:
The patent segments the classification process into two stages: initial filtering by the first classifier and detailed classification by the second classifier. This segmentation reduces the computational processing power required by the second classifier, as it only needs to process segments that were not confidently classified by the first stage, while still achieving complete classification for all segments.
Solution Approach 2:
The first stage classifier performs preliminary filtering to identify and eliminate segments that do not require further classification. This preliminary action reduces the workload for the second stage classifier, lowering the overall computational processing power requirements while ensuring that all segments requiring classification receive complete classification.
Data Source
AI summary
A system and method is disclosed for classifying time-series data provided to a machine-learning model from a continuous sensor signal. The data may be “windowed” or “divided” into a smaller data segment using a first stage classifier where an “event of interest” may be identified. The first stage classifier may employ an algorithm that prohibits false negative identifications. The data segment detected as including an event of interest may then be transmitted to a second stage classifier operable to performs a full classification on the data segment. The multi-stage network may require less power and a less complex structure.