Server-Sensor Data Collection for AI Training Imbalance
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning technologies face challenges in collecting appropriate data for training models, leading to failures in achieving desired performance, particularly when using unsuitable data, resulting in imbalanced datasets that can lead to unfairness in recognition rates for minority attributes like children in image recognition systems.
Innovation Solution
A data collection system comprising a sensor device and a server device that collaboratively collect and analyze data, where the server device identifies beneficial or lacking data for training using Explainable AI or influence functions, and the sensor device collects and transmits data to the server for retraining, ensuring data quality and privacy compliance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If data is collected without analysis of training needs, then data collection is simple, but the learned model performance is poor
Solution Approach 1:
The system uses a data analysis unit to analyze the learning model's training status and provide feedback information about which data is beneficial or lacking. This feedback loop enables the system to collect appropriate data based on actual training needs, improving model performance while maintaining targeted data collection
Solution Approach 2:
The learning model itself provides information about its training needs through the data analysis unit, which identifies beneficial and lacking data. The system serves itself by using the model's own performance characteristics to guide data collection, reducing the need for external intervention
2Reliability
If inappropriate data is used for training, then data collection is easy, but the model fails to achieve desired performance
Solution Approach 1:
The data analysis unit continuously analyzes the learning model and provides feedback on data quality and suitability. This feedback mechanism ensures that only appropriate data is collected and used for training, preventing the use of inappropriate data while maintaining efficient data collection
3Measurement precision
If data is collected to improve minority attribute recognition, then recognition accuracy improves, but data imbalance may worsen without proper analysis
Solution Approach 1:
The data analysis unit analyzes the learning model's performance across different attributes and provides feedback on data distribution. This enables the system to identify which minority attributes need more data and collect data in proportions that improve recognition accuracy while maintaining overall data balance
Solution Approach 2:
The system applies different data collection strategies to different data types based on specific needs. For minority attributes with insufficient data, the system collects more samples, while for well-represented attributes, collection is reduced or stopped, achieving local optimization of data quality
Data Source
AI summary
A data collection system according to the present disclosure includes: a sensor device that collects data; anda server device including a learning model that performs output according to a learning result, corresponding to input, and a data analysis unit that specifies data that is beneficial for or lacking in training of the learning model. The server device transmits, to the sensor device, a request signal for collecting data that is beneficial for or lacking in the training specified by the data analysis unit, or data similar to the data, the sensor device collects data that is beneficial for or lacking in the training, or similar data based on the received request signal, and transmits the collected data to the server device, and the server device retrains the learning model based on the data transmitted from the sensor device.