Neural Network Object-Class Model Parameter Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional AI models based on deep learning require extensive training time, making it difficult for them to quickly adapt to changes in data distributions, such as those encountered in real-time production scenarios, where sudden demands for defect detection in manufacturing cannot be met within the required short timeframe.
Innovation Solution
A method and device for optimizing an object-class model based on a neural network that allows for rapid adjustment and training of parameters when new target images are encountered, enabling the model to correctly classify new images by selecting and modifying specific parameters, thereby accommodating new data without significantly altering the existing model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional deep learning models are trained on single GPU, then model training can be performed, but training time becomes excessively long (3 hours or longer)
Solution Approach 1:
The patent segments the model training process into two distinct phases: (1) initial comprehensive training on single GPU to establish baseline model, and (2) incremental fine-tuning training that only updates selected parameters when new defect types are encountered. This segmentation allows the system to avoid full retraining while adapting to new data distributions, reducing training time from hours to minutes.
Solution Approach 2:
The patent implements dynamic training strategy where the training process adapts based on the detected needs. When new defect types are identified, the system dynamically adjusts which parameters are updated and for how long, rather than following a fixed training schedule. This dynamic approach optimizes the balance between model adaptability and training time consumption.
2Adaptability or versatility
If model parameters are extensively updated to accommodate new data distributions, then model adaptability improves, but training time increases significantly
Solution Approach 1:
The patent extracts and updates only the necessary parameters when new defect types are encountered, rather than retraining the entire model. Specifically, it identifies and fine-tunes parameters related to new defect characteristics while keeping other parameters fixed, thus achieving adaptability with minimal training time investment.
Solution Approach 2:
The patent applies local quality by differentiating the treatment of model parameters based on their relevance to new defect types. Parameters directly related to new defect characteristics receive updated training, while parameters unrelated to new defects remain unchanged. This localized parameter updating achieves adaptability without the time cost of global retraining.
3Measurement precision
If full model retraining is performed to handle sudden production demands, then classification accuracy improves, but production downtime increases
Solution Approach 1:
The patent performs preliminary action by pre-training the model comprehensively on single GPU to establish a robust baseline that handles most defect types. This preliminary training creates a model that requires minimal subsequent updates, allowing the system to respond quickly to sudden production demands without extensive retraining, thus maintaining production continuity while preserving detection accuracy.
Data Source
AI summary
The present disclosure discloses a method and device for optimizing an object-class model based on a neural network. The method includes: establishing the object-class model based on the neural network, training the object-class model, and realizing classification of target images by using the object-class model that has been trained; and when a new target image is generated, and the new target image is an image corresponding to a new condition of a target and is capable of still being classified into an original classification system, judging a result of identification of the object-class model to the new target image, and if the object-class model is not capable of correctly classifying the new target image, according to the new target image, selecting some of parameters, adjusting the some of parameters, and training to obtain an object-class model that is capable of correctly classifying the new target image.