Distributed Subclassification Model for Edge Terminals
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep learning classification models face accuracy loss and computational inefficiencies when executed on low-performance devices like smartphones and IoT terminals, as techniques like pruning reduce model weights but decrease accuracy, and distributed processing methods are not stable.
Innovation Solution
A classification method employing distributed subclassification models, where pretrained models are allocated to multiple terminals based on resource monitoring, with each model handling a subset of target classes, allowing for stable and accurate classification even if some terminals fail.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Volume of moving object
If pruning technique is used to reduce model size, then model size is reduced, but classification accuracy significantly decreases
Solution Approach 1:
The patent divides a large classification model into multiple smaller subclassification models, each handling a subset of classes. This segmentation allows each model to be lightweight while collectively maintaining high classification accuracy across all classes, resolving the contradiction between model size reduction and accuracy preservation.
2Measurement precision
If deep learning model complexity is increased to improve accuracy, then classification accuracy improves, but computational load and resource requirements increase
Solution Approach 1:
By segmenting the classification task into multiple specialized subclassification models, each model can be simpler and less computationally intensive while collectively achieving high accuracy. This divides the computational load across multiple models rather than requiring one complex model.
Solution Approach 2:
Each subclassification model is optimized for specific classes it handles, allowing local optimization rather than requiring all models to be universally complex. This enables simpler, more efficient models for each local classification task.
3Power
If distributed processing is implemented across multiple terminals, then computational load is distributed, but system stability decreases due to potential terminal failures
Solution Approach 1:
The patent changes the parameter of model distribution by allocating different subclassification models to different terminals based on their capabilities. This parameter change enables effective distributed processing while maintaining system stability through adaptive allocation.
4Adaptability or versatility
If low-performance devices are used for deep learning, then device accessibility improves, but model execution performance deteriorates
Solution Approach 1:
Segmenting the classification model into smaller subclassification models enables deployment on low-performance devices that cannot handle large complex models. Each small model can execute efficiently on resource-constrained devices while maintaining overall system accuracy.
Solution Approach 2:
The patent changes the model parameters by creating lightweight subclassification models with fewer weights and parameters, making them suitable for execution on low-performance devices while maintaining classification effectiveness through the collective ensemble of models.
Data Source
AI summary
Disclosed is a method for classifying target data into a particular class by distributing a previously learned classification model to a plurality of terminals. A disclosed classification method using a distributed classification model comprises the steps of: receiving classification data which is for target data and generated by means of subclassification models respectively distributed to a plurality of terminals; and determining a final class for the target data by means of the classification data. Each class for classification allocated to the previously learned subclassification models comprises at least one of a plurality of target classes. The number of classes allocated to the respective subclassification models is less than the number of the target classes.