Distributed Training Master Node Detects Deviation Parameters
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Inaccurate data contamination during AI model training in distributed systems leads to incorrect judgments and production issues, as mislabeled or irrelevant data can significantly impact the accuracy of AI models deployed in industrial settings like production lines.
Innovation Solution
A method for distributed training that identifies contaminated data by determining deviation parameters through the distribution of local model parameters, using a master computing apparatus to update the global model parameters, thereby improving the accuracy of the AI model by excluding contaminated data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If distributed training is performed using data from multiple production lines, then the AI model can be trained with diverse data to improve generalization, but contaminated data from any production line can affect the accuracy of the trained model
Solution Approach 1:
The system implements a feedback mechanism where the master computing apparatus monitors the training process and compares local model parameters against global parameters. When deviation parameters are detected, the system feeds back to identify and exclude contaminated data from the training set, thereby maintaining model accuracy while preserving the benefits of diverse data from multiple production lines
Solution Approach 2:
The master computing apparatus acts as an intermediary between multiple worker computing apparatuses. It collects local model parameters from all production lines, identifies deviation parameters that indicate contamination, and mediates the training process by excluding contaminated data while incorporating clean data from other production lines, thus resolving the contradiction between diversity and accuracy
2Productivity
If all local model parameters are used to update the global parameter, then the training process is simple and efficient, but contaminated data can propagate and degrade model performance
Solution Approach 1:
The system extracts and removes contaminated data (deviation parameters) from the training process. The master computing apparatus identifies deviation parameters through distribution comparison and excludes them from the global parameter update, thereby maintaining training efficiency while preventing contamination from degrading model performance
Solution Approach 2:
The system changes the parameter update mechanism by introducing a filtering step. Instead of uniformly applying all local parameters, the system modifies the parameter selection process to exclude deviation parameters, thus changing the effective parameters used in global model updates to maintain both efficiency and accuracy
3Adaptability or versatility
If data from multiple production lines is collected for training, then the AI model can handle diverse production scenarios, but mislabeled or irrelevant data can be inserted and cause data contamination
Solution Approach 1:
The system implements continuous feedback during the training process where the master computing apparatus monitors local model parameters for deviation from the global distribution. When contamination is detected through statistical comparison, the system feeds back to identify and remove contaminated data, thereby maintaining the ability to handle diverse production scenarios while preventing harmful data from degrading model performance
Solution Approach 2:
The master computing apparatus serves as an intermediary that collects data from multiple production lines, identifies contamination through distribution comparison, and filters out harmful data while preserving useful data from diverse scenarios. This intermediary role enables the system to maintain adaptability across different production lines while eliminating the harmful effects of mislabeled or irrelevant data
Data Source
AI summary
An optimizing method of distributed training and a master computing apparatus are provided. In the method, a local model is trained by using one of multiple sample sets and a global parameter of a global model to generate a local parameter of the local model. One or more deviation parameters among the local parameter of multiple local models trained by the sample sets are determined. A distribution of the deviation parameter is far from a distribution of other local parameters, and the local parameter of the local models is used to update the global parameter of the global model. Accordingly, the prediction accuracy of the global model may be improved.