Feature Extraction Model Training Without Labeled Teacher Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing feature extraction models require teacher data for learning, which is difficult to prepare, especially in manufacturing processes like semiconductor device defect detection, leading to suboptimal inference accuracy.
Innovation Solution
An information processing apparatus that selects analysis target and hint data without labeled teacher data, constructs a machine learning model to extract features, calculates indices for feature separation and extraction, and updates model weights based on these indices, using techniques like convolutional neural networks and vision transformers.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If teacher data is used for model learning, then feature extraction accuracy is improved, but data preparation difficulty increases
Solution Approach 1:
The model performs self-supervised learning by utilizing the input data itself as the teaching signal. The system automatically generates pseudo-labels and features without requiring external teacher data, enabling the model to learn and improve through self-service rather than relying on manually prepared teacher datasets.
Solution Approach 2:
An intermediate feature extraction process is introduced that generates pseudo-labels and intermediate features from the input data itself. These intermediates serve as substitutes for traditional teacher data, allowing the model to learn through self-generated signals rather than requiring external labeled datasets.
2Reliability
If labeled teacher data is required for learning, then model learning effectiveness is improved, but learning efficiency deteriorates
Solution Approach 1:
The system eliminates the need for external teacher data by using self-supervised learning mechanisms. The model automatically generates learning signals from the input data itself, removing the time-consuming process of data collection, labeling, and verification while maintaining learning effectiveness through self-generated pseudo-labels and features.
Solution Approach 2:
The system performs preliminary feature extraction and pseudo-label generation automatically during the learning process itself, rather than requiring these to be prepared in advance as separate steps. This eliminates preprocessing bottlenecks and allows the model to learn directly from self-generated signals.
3Ease of manufacture
If teacher data preparation is simplified, then ease of manufacture is improved, but feature extraction accuracy deteriorates
Solution Approach 1:
The model generates high-quality pseudo-labels and intermediate features automatically through self-supervised learning mechanisms. These self-generated signals maintain or even improve feature extraction accuracy compared to traditional teacher data, while eliminating the need for manual data preparation entirely.
Solution Approach 2:
The system changes the fundamental parameter of data labeling from manual human annotation to automated algorithmic generation. This parameter change enables both ease of data preparation and maintains accuracy by using learned features from the model itself rather than simplified but potentially inaccurate manual labels.
Data Source
AI summary
An information processing apparatus comprising processing circuitry configured to select analysis target data, select hint data related to a feature to be noted, construct a machine learning model that extracts a first feature included in the analysis target data and a second feature included in the hint data based on the selected analysis target data and the selected hint data, calculate a first index for evaluating the first feature, calculate a second index for evaluating the second feature, and update a weight of the machine learning model based on the first index and the second index.