Deep Learning Diagnostic System for Semiconductor Image Analysis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current quality assurance methods for deep learning models in semiconductor applications cannot identify correct predictions based on wrong causal features and treat machine learning algorithms as black boxes, making it difficult to improve poorly trained classifiers.
Innovation Solution
A system and method for performing diagnostic functions on deep learning models, which includes a deep learning model and a diagnostic component to determine causal portions of images and perform functions based on those determinations, providing causal understanding and guidance for data augmentation and fine-tuning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning models are used for defect classification, then classification accuracy is improved, but interpretability and diagnostic capability deteriorate due to black box operation
Solution Approach 1:
The patent introduces gradient-based visualization techniques as intermediaries between the deep learning model and the user. By computing gradients of the loss function with respect to input pixels and visualizing them, the system creates a bridge that reveals which input regions most influenced the classification decision, thus providing interpretability without sacrificing the model's accurate classification capability
Solution Approach 2:
The patent replaces traditional mechanical quality assurance methods (manual inspection, rule-based classification) with deep learning models that use gradient computations and visualization techniques. This substitution enables automated diagnostic capabilities that can identify causal features while maintaining high classification accuracy, overcoming the black box limitation
2Device complexity
If currently used quality assurance approaches are applied, then model evaluation is simplified, but the ability to identify correct predictions based on wrong causal features is lost
Solution Approach 1:
The patent introduces gradient visualization as an intermediary diagnostic tool that computes and displays the spatial distribution of gradients for each class. This intermediary mechanism enables precise identification of which input features drove the classification, allowing detection of cases where the model made correct predictions based on incorrect causal features, without significantly increasing evaluation complexity
Solution Approach 2:
The patent uses color-coded gradient visualizations to represent different classes and their influence on the classification. By displaying gradient magnitudes and directions through color variations, the system enables intuitive visual diagnosis of classification decisions, allowing practitioners to identify erroneous causal feature usage while maintaining simple evaluation procedures
3Adaptability or versatility
If data augmentation is performed without causal understanding, then training data diversity is improved, but the ability to directly improve poorly trained classifiers is reduced
Solution Approach 1:
The patent implements a feedback mechanism where gradient visualizations from the trained model inform subsequent data augmentation decisions. By analyzing which features the model relies on (through gradient patterns), practitioners can generate augmented data that specifically targets weak or erroneous feature associations, creating a feedback loop that continuously improves classifier performance based on diagnostic insights
Solution Approach 2:
The patent enables preliminary diagnostic analysis using gradient visualizations before performing data augmentation. This preliminary action identifies which classes or features require improved training, allowing practitioners to针对性地 generate augmented data for specific weak points rather than applying generic augmentation, thereby directly improving poorly trained classifiers
Data Source
AI summary
Methods and systems for performing diagnostic functions for a deep learning model are provided. One system includes one or more components executed by one or more computer subsystems. The one or more components include a deep learning model configured for determining information from an image generated for a specimen by an imaging tool. The one or more components also include a diagnostic component configured for determining one or more causal portions of the image that resulted in the information being determined and for performing one or more functions based on the determined one or more causal portions of the image.