NLP Classifier Continual Learning with Synthetic Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing natural language processing (NLP) models struggle to learn new classes without degrading accuracy on existing classes, requiring large amounts of annotated training data and storage of original datasets, which poses data security risks and resource inefficiencies.
Innovation Solution
A continual few-shot learning approach generates synthetic training data from the NLP model to represent existing classes, allowing the model to be retrained with minimal new data, thus preserving accuracy on old classes and reducing resource and storage needs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If the model is retrained with new class data added to the original large-scale dataset, then the model can recognize new classes, but the storage requirements and data security risks increase due to preserving original training data
Solution Approach 1:
The patent creates synthetic training data that copies the essential characteristics of original training data through model-generated synthetic samples. This allows the model to learn new classes without storing the original large-scale training dataset, reducing storage requirements while maintaining the ability to recognize both old and new classes
Solution Approach 2:
The patent changes the parameters of training data by generating synthetic data with modified characteristics that preserve the semantic meaning of original data. This transformation allows the model to learn from compressed representations rather than raw data, reducing storage needs while maintaining adaptability to new classes
2Measurement precision
If a large number of annotated training instances are generated for the new class, then the model accuracy for new classes improves, but the time and resources required for data annotation increase
Solution Approach 1:
The patent enables the model to generate its own training data through synthetic data generation. The model uses its current knowledge to create synthetic examples of new classes, eliminating the need for manual annotation by human experts and significantly reducing the time and resources required while maintaining high accuracy
Solution Approach 2:
The patent performs preliminary synthetic data generation before the actual training process. By pre-generating synthetic training instances that capture the essential characteristics of new classes, the model is prepared for efficient training without requiring time-consuming manual annotation during the training phase
3Adaptability or versatility
If the model is trained on new classes, then the model can classify new data classes, but the accuracy on old classes degrades due to catastrophic forgetting
Solution Approach 1:
The patent introduces synthetic data as an intermediary between the model's existing knowledge and new class information. The synthetic data acts as a bridge that preserves the characteristics of old classes while incorporating new class information, allowing the model to learn new classes without forgetting old ones
Solution Approach 2:
The patent maintains continuous learning by generating synthetic data that incorporates both old and new class information throughout the training process. This continuous exposure to both old and new class characteristics through synthetic data prevents catastrophic forgetting while enabling the model to acquire new classification abilities
Data Source
AI summary
Embodiments of the technology described herein describe a machine classifier capable of continually learning new classes through a continual few-shot learning approach. A natural language processing (NLP) machine classifier may initially be trained to identify a plurality of other classes through a conventional training process. In order to learn a new class, natural-language training data for a new class is generated. The training data for the new class may be few-shot training data. The training also uses synthetic training data that represents each of the plurality of other classes. The synthetic training data may be generated through a model inversion of the original classifier. The synthetic training data and the natural-language training data are used to retrain the NLP classifier to identify text in the plurality of other classes and the new class using.