Generalized Balanced Few-Shot Learning Without Base Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing generalized few-shot learning methods require base data, which may not be available due to privacy or ethical issues, and fail to properly balance the mean and variance of weight distributions between novel and base classes, leading to unsatisfactory performance.
Innovation Solution
A method and system for generalized balanced few-shot learning that pre-trains a feature extractor and classifier using base data, then fine-tunes a joint linear classifier while freezing the feature extractor, performing weight normalization to achieve zero-mean and balanced variance, allowing knowledge from novel data to be incorporated into a pre-trained model without base data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If transfer learning is used to quickly adapt to novel classes, then learning speed is improved, but performance on base classes deteriorates due to catastrophic forgetting
Solution Approach 1:
The classifier weights are segmented into two distinct parts: a frozen pre-trained classifier for base classes and a trainable novel classifier for novel classes. This segmentation allows independent optimization of each classifier, preventing interference between base and novel class learning while maintaining both capabilities simultaneously
Solution Approach 2:
A projection layer is introduced as an intermediary component that maps features from the frozen feature extractor to the trainable novel classifier. This intermediary enables knowledge transfer to novel classes without requiring updates to the pre-trained model, thus preserving base class performance while achieving fast adaptation
2Reliability
If additional architecture is added to preserve base class knowledge, then performance on base classes is maintained, but device complexity increases
Solution Approach 1:
The frozen pre-trained model serves multiple functions simultaneously: it acts as both the feature extractor and the base class classifier. By making the pre-trained model multi-functional, the need for separate dedicated components is reduced, simplifying the overall architecture while maintaining base class performance
Solution Approach 2:
The novel class learning capability is extracted as a separate, lightweight trainable classifier that operates independently from the frozen pre-trained model. This extraction allows the system to gain novel class adaptability without modifying or complicating the proven base class classification architecture
3Reliability
If base data is used for pre-training, then performance on novel classes is improved, but the method becomes inapplicable when base data is unavailable due to privacy or ethical issues
Solution Approach 1:
Instead of requiring access to original base data, the method uses a copy of base data in the form of pre-trained model weights. The pre-trained model encapsulates knowledge from base classes, allowing the system to benefit from base class information without needing actual access to the sensitive base data itself
Solution Approach 2:
The pre-training process is performed in advance on base data to create a pre-trained model that already contains base class knowledge. This preliminary action enables the model to be deployed in scenarios where base data is unavailable, as the knowledge has already been encoded into the model weights before deployment
4Reliability
If weight normalization is performed to achieve zero-mean and balanced variance, then balanced inference between base and novel classes is achieved, but computational steps increase
Solution Approach 1:
Weight normalization transforms the weight parameters of the novel classifier by adjusting their mean to zero and their variance to match that of the base classifier. This parameter transformation creates a balanced weight distribution that prevents any single class from dominating the decision boundary, achieving equitable performance across all classes
Data Source
AI summary
A method and system of generalized balanced few-shot learning using only novel data without old data is proposed. The method of generalized balanced few-shot learning using only novel data without old data proposed in the present disclosure includes pre-training stage for training a feature extractor and a classifier of a training model with base data through a pre-training unit and fine-tuning stage for freezing the feature extractor through a fine-tuning unit, training a joint linear classifier capable of inferring base classes and novel classes, and performing weight normalization to achieve zero-mean and balanced variance.