Classification System Using Shapley Value Distillation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine learning methods for classifying defects in display panel manufacturing face challenges in accurately identifying repairable and unrepairable defects using sensory data, particularly due to difficulties in capturing small defects with computer vision and the need for costly incremental updates in decision-tree based models, which are not easily adaptable to data drift.
Innovation Solution
A classification system that uses multi-task distillation to train a neural network based model with Shapley values, obtained from a decision-tree based model, to predict class labels and provide explanations, allowing for incremental updates and adaptation to new data without requiring all historical data, thereby reducing computational costs and complexity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If decision-tree based models are used for classification, then interpretability and ease of understanding are improved, but adaptability to data drift and incremental updates are worsened
Solution Approach 1:
The patent uses Shapley values as an intermediary to bridge decision-tree models and neural networks. The decision-tree model provides interpretable Shapley values that serve as guidance signals to train the neural network, enabling the neural network to learn from the interpretable rules while maintaining its own predictive capabilities and adaptability to data drift.
Solution Approach 2:
The patent copies the knowledge and decision-making patterns from the decision-tree model into the neural network through knowledge distillation. By training the neural network to replicate the Shapley value predictions of the decision-tree model, the neural network inherits the interpretability benefits while gaining the flexibility and adaptability of neural network architectures.
2Adaptability or versatility
If neural network models are used for classification, then adaptability and learning capability are improved, but interpretability and computational complexity are worsened
Solution Approach 1:
Shapley values serve as an intermediary that simplifies the interpretation of neural network predictions. By calculating and analyzing Shapley values for the neural network model, the patent provides a systematic method to understand feature importance and model decisions, reducing the interpretability complexity of neural networks.
Solution Approach 2:
The patent performs preliminary action by pre-calculating Shapley values from a decision-tree model and using them as training targets for the neural network. This preliminary computation of feature importance guides the neural network training process, reducing the overall computational complexity by providing directed learning signals rather than requiring complex post-hoc analysis.
3Adaptability or versatility
If incremental updates are performed on decision-tree models, then adaptability to new data is improved, but computational costs and complexity are worsened
Solution Approach 1:
The patent copies the adaptive capability from decision-tree models to neural networks through knowledge distillation. Instead of performing incremental updates on complex decision-tree structures, the system trains the neural network to replicate the Shapley value predictions, enabling efficient incremental adaptation by simply retraining the neural network on new data with updated Shapley value targets.
Solution Approach 2:
The patent uses a cost-effective approach by training relatively simple neural network models that can be quickly retrained on new data. These neural networks serve as disposable or easily replaceable components that can be incrementally updated without the high computational costs associated with maintaining and updating complex decision-tree models, especially for capturing small defects in manufacturing data.
Data Source
AI summary
A classification system includes: one or more processors; and memory including instructions that, when executed by the one or more processors, cause the one or more processors to: calculate reference Shapley values for features of a data sample based on a first classification model; and train a second classification model though multi-task distillation to: predict Shapley values for the features of the data sample based on the reference Shapley values and a distillation loss; and predict a class label for the data sample based on the predicted Shapley values and a ground truth class label for the data sample.