AI Server Classification Layer Update for Ethnic Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing AI models struggle to maintain inference accuracy across different ethnic groups, requiring retraining when switching from white-oriented to Asian-oriented data, and often fail to meet user requirements for general-purpose models.
Innovation Solution
An AI server capable of updating and providing a classification layer, which includes training the AI model with classification training data and labeling data to optimize the classification layer for specific ethnic groups, while keeping the feature extraction layer unchanged, thereby reducing the need for full model retraining.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the entire AI model is retrained with Asian-oriented training data, then the inference accuracy for Asians is improved, but the training cost and time increase significantly
Solution Approach 1:
The patent segments the AI model into two independent parts: a feature extraction layer and a classification layer. By updating only the classification layer with Asian-oriented training data while keeping the feature extraction layer unchanged, the system achieves improved inference accuracy for Asians without the need to retrain the entire model, thereby significantly reducing training time and computational resources.
Solution Approach 2:
The patent extracts and updates only the classification layer from the complete AI model. This selective extraction allows the system to focus training resources on the specific component responsible for classification decisions, improving efficiency while maintaining the benefits of the pre-trained feature extraction layer.
2Measurement precision
If the entire AI model is retrained with Asian-oriented training data, then the inference accuracy for Asians is improved, but the training cost increases significantly
Solution Approach 1:
The patent segments the AI model into two independent parts: a feature extraction layer and a classification layer. By updating only the classification layer with Asian-oriented training data while keeping the feature extraction layer unchanged, the system achieves improved inference accuracy for Asians without the need to retrain the entire model, thereby significantly reducing training time and computational resources.
Solution Approach 2:
The patent applies partial action by updating only the necessary classification layer rather than the entire model. This approach uses just enough training resources to achieve the desired improvement in inference accuracy for Asian faces, avoiding the excessive computational cost of full model retraining.
3Adaptability or versatility
If a general-purpose AI model is used for both whites and Asians, then the model can be applied universally, but the performance does not meet user requirements for specific ethnic groups
Solution Approach 1:
The patent implements a dynamic model architecture where the classification layer can be flexibly updated for different ethnic groups while maintaining the universal feature extraction layer. This allows the system to adapt to specific user requirements (e.g., Asian-oriented or white-oriented classification) by swapping or retraining only the classification layer, thereby achieving both universality and specialized performance.
Solution Approach 2:
The patent applies local quality by optimizing the classification layer for specific ethnic groups while keeping the feature extraction layer general-purpose. This allows different parts of the model to have different specialization levels, with the classification layer tailored to local requirements (Asian or white) and the feature extraction layer maintaining universal applicability.
Data Source
AI summary
An artificial intelligence (AI) server is provided. The AI server includes a communication interface configured to communicate with an electronic device, and at least one processor configured to update a classification layer by training an artificial intelligence model in such a manner that classification training data and classification labeling data are provided to the artificial intelligence model including a feature extraction layer for extracting a feature vector and a classification layer for classifying input data using the feature vector, and transmit the updated classification layer to the electronic device.