Server-Side Image Recognition Model Evaluation for Unsupervised Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional image recognition using supervised data requires costly preparation and deteriorates when faced with data having a different distribution, leading to suboptimal recognition performance.
Innovation Solution
An information processing device and method that employs multiple recognition units updated through unsupervised learning across terminals, evaluating recognition results to determine correct learning and adjusting models accordingly, allowing image recognition without relying on correct answer labels.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If supervised learning is used for image recognition, then recognition accuracy can be improved with sufficient training data, but the cost of preparing supervised data increases and recognition performance deteriorates when data distribution differs from training data
Solution Approach 1:
The system enables terminals to perform unsupervised learning on their own local data without requiring external supervised labels. Each terminal autonomously updates its recognition model using only local image data, eliminating the need for centralized supervised data preparation and labeling operations.
Solution Approach 2:
Instead of using expensive supervised data with correct answer labels, the system copies and utilizes the natural data distribution from terminals. The unsupervised learning process learns from the inherent patterns and structures in the data itself, rather than requiring人工-annotated ground truth data.
2Measurement precision
If supervised learning is used for image recognition, then recognition accuracy can be improved, but recognition performance deteriorates when input data has different distribution from training data
Solution Approach 1:
The system dynamically adapts the recognition model to local data distributions through unsupervised learning. Instead of using a fixed supervised training approach, the model continuously adjusts its parameters based on the actual data characteristics encountered during inference, enabling it to handle varying data distributions effectively.
Solution Approach 2:
The unsupervised learning process modifies model parameters adaptively based on the statistical properties of input data. The system changes parameter representations and learning dynamics to match the specific data distribution at each terminal, improving generalization across different environments without requiring retraining.
3Measurement precision
If multiple recognition units with updated models are deployed, then recognition performance on diverse data can be improved, but device complexity increases
Solution Approach 1:
The system segments the recognition task across multiple independent terminal devices, each maintaining its own recognition unit. This distributed architecture allows different terminals to specialize in their local data distributions while sharing the overall system objective, reducing the complexity burden on any single centralized system.
Solution Approach 2:
Each recognition unit is designed with universal functionality to handle various data types and distributions through unsupervised learning. The same basic recognition unit architecture can operate effectively across different terminals with different data characteristics, eliminating the need for specialized models for each scenario.
Data Source
AI summary
An information processing device on a server side includes: a predetermined number of recognition units, for which a model updated by performing image recognition in a predetermined number of vehicles and executing unsupervised learning is each set, configured to perform image recognition on an image, on which image recognition has been performed in a predetermined number of the vehicles; and an evaluation value calculation unit configured to evaluate recognition results obtained in a predetermined number of the recognition units and calculate an evaluation value for each of the recognition units. The information processing device on the vehicle side includes an execution unit that executes unsupervised learning, and a determination unit that determines whether learning has been performed correctly or not for a model updated in the execution unit on the basis of an evaluation value found on the server side.