Category Mapping Discriminators for Few-Shot Image Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional deep neural networks (DNNs) face challenges in incremental learning with few training samples, experiencing overfitting and catastrophic forgetting, which limits their ability to classify new categories efficiently and accurately, unlike human learning which can generalize quickly with limited examples.
Innovation Solution
A method and system for image classification using category mapping discriminators, where a set of discriminators are configured to recognize features of input images, allowing for the generation of new discriminators without retraining existing ones, enabling fast and effective few-shot incremental learning by mapping original features to respective categories without losing discriminative characteristics.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep neural networks are trained using standard supervised learning with gradient-based approaches, then high classification accuracy is achieved, but a large amount of training samples are required and retraining is needed when new categories are added
Solution Approach 1:
The classification task is segmented into two independent components: (1) a pre-trained feature extractor that captures general visual features, and (2) category-specific mapping discriminators that map features to category embeddings. This segmentation allows the feature extractor to be trained on large datasets once, while new categories are added by training only their specific discriminators on few samples, resolving the contradiction between accuracy and sample quantity.
Solution Approach 2:
The feature extractor is pre-trained on a large source dataset before being used for target category classification. This preliminary action on the source domain enables the model to learn generalizable features that can be transferred to target categories with limited samples, achieving high accuracy without requiring large amounts of target training data.
2Adaptability or versatility
If deep neural networks are used for incremental class learning, then new categories can be learned, but catastrophic forgetting of previously learned categories occurs
Solution Approach 1:
The network is segmented into a shared feature extractor and category-specific mapping discriminators. When new categories are added, only the new discriminators are trained while the feature extractor remains frozen. This segmentation prevents interference with previously learned categories, eliminating catastrophic forgetting while enabling incremental learning of new classes.
Solution Approach 2:
Category embeddings serve as an intermediary representation that decouples the feature extraction process from category-specific classification. The mapping discriminators learn to map features to these fixed embeddings without modifying the feature extractor, allowing new categories to be added without retraining or forgetting old categories.
3Device complexity
If conventional few-shot learning methods using embedding and metric learning are applied, then computational complexity is reduced, but performance relies on the assumption that the embedding is sufficiently discriminative which may not hold for new tasks
Solution Approach 1:
Instead of using fixed linear comparators, the patent employs dynamic mapping discriminators that are trained to adaptively learn optimal non-linear mappings from features to category embeddings. This dynamic adaptation enables the system to achieve high discriminative capability for new tasks while maintaining computational efficiency through the use of pre-extracted features.
4Adaptability or versatility
If optimizer learning methods like MAML or Meta-SGD are used for few-shot learning, then fine-tuning strategies are learned, but fine-tuning is required which increases computational complexity
Solution Approach 1:
The patent extracts and freezes the general feature extraction capability from a pre-trained network, separating it from the category-specific classification task. This extraction allows new categories to be classified by training only lightweight mapping discriminators without requiring fine-tuning of the entire network, thereby achieving few-shot adaptability with reduced computational complexity.
Data Source
AI summary
There is provided a method of image classification. The method includes: providing a set of category mapping discriminators, each corresponding to a respective category, wherein each category mapping discriminator of the set of category mapping discriminators is configured for discriminating features relating to input images that belong to the respective category of the category mapping discriminator; extracting a plurality of features from an input image using a machine learning model; determining, for each of the set of category mapping discriminators, an output value based on the plurality of extracted features using the category mapping discriminator; and determining a classification of the input image based on the output values of the set of category mapping discriminators.