Recognition Model Integration via Linear Summation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing recognition models require extensive training time and resources when updating with large amounts of additional data, as they need to retrain from scratch, which is inefficient and resource-intensive.
Innovation Solution
A device and method that integrates a first recognition model with a second model generated from additional training data using a linear summation approach, allowing for efficient updating of recognition models by selecting weights based on the quantity and importance of data, thereby reducing the need for extensive retraining.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the model is trained with the sum of prior existing training data and additional training data, then the model can learn from all available data, but the training time becomes excessively long
Solution Approach 1:
The patent segments the training process into two distinct phases: (1) initial model training using only basic training data, and (2) subsequent fine-tuning using only additional training data. This segmentation avoids the need to retrain on the complete dataset (basic + additional), thereby significantly reducing training time while maintaining model learning effectiveness.
Solution Approach 2:
The patent performs preliminary training action by first training the model on basic training data to establish an initial recognition model. This preliminary model serves as a foundation that can be quickly adapted to additional data without requiring complete retraining, thus reducing overall training time while ensuring the model learns from all available data.
2Reliability
If the model is trained with large amounts of data, then the model performance improves, but the computational resources required increase
Solution Approach 1:
The training process is segmented into two efficient stages: initial training on basic data to establish performance foundation, and subsequent fine-tuning on additional data to improve performance further. This segmentation avoids the computationally expensive process of training from scratch on the complete large dataset, thereby reducing computational resource requirements while maintaining model performance improvement.
Solution Approach 2:
The patent changes training parameters by using different learning rates or optimization settings for the two training stages. The initial training uses parameters optimized for learning from basic data, while the fine-tuning stage uses parameters optimized for adapting to additional data. This parameter adaptation allows the model to achieve performance improvement from large amounts of data with reduced computational overhead compared to uniform training approaches.
3Adaptability or versatility
If the model is retrained from scratch with all training data, then the model can incorporate new data, but the process is inefficient and resource-intensive
Solution Approach 1:
The patent performs preliminary training action by first establishing a recognition model using basic training data. This pre-trained model serves as a starting point that can be efficiently updated with additional data through fine-tuning, rather than requiring complete retraining from scratch. This approach maintains model adaptability to new data while dramatically improving updating efficiency.
Solution Approach 2:
The patent changes training parameters between the initial model creation and subsequent updates. By using different optimization parameters, learning rates, or training configurations for the initial training versus the update phase, the system achieves efficient model updates that maintain adaptability without the inefficiency of complete retraining, thereby improving productivity in model updating operations.
Data Source
AI summary
A method and a device for recognition, and a method and a device for constructing a recognition model are disclosed. A device for constructing a recognition model includes a training data inputter configured to receive additional training data, a model learner configured to train a first recognition model constructed based on basic training data to learn the additional training data, and a model constructor configured to construct a final recognition model by integrating the first recognition model with a second recognition model generated by the training of the first recognition model.