Meta Few-Shot Class Incremental Learning via Attention Maps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Classic machine learning algorithms face challenges in adapting to new classes without forgetting old knowledge, especially with limited training samples, leading to catastrophic forgetting and overfitting issues in incremental learning scenarios.
Innovation Solution
The method involves generating weight and feature attention maps to update model weights and feature maps, allowing for a balance between learning new knowledge and maintaining old knowledge through a modulatory network and prediction network interaction, with a bi-directional guided adaptation mechanism.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If the model is trained to learn new classes, then the adaptability to new knowledge is improved, but the old knowledge is forgotten (catastrophic forgetting)
Solution Approach 1:
The model is divided into two independent networks: a prediction network that maintains old knowledge and a modulatory network that learns new knowledge. This segmentation allows each network to specialize in different tasks, preventing the modulatory network from interfering with the prediction network's learned knowledge, thus resolving catastrophic forgetting while enabling continuous adaptation to new classes.
Solution Approach 2:
The modulatory network acts as an intermediary between new input data and the prediction network. It processes new classes and generates attention maps that guide the prediction network's weight updates, allowing the model to adapt to new knowledge without directly modifying the prediction network's core parameters, thereby preventing forgetting of old knowledge.
2Adaptability or versatility
If the model updates weights to adapt to new classes, then the learning of new knowledge is improved, but the stability of old knowledge deteriorates
Solution Approach 1:
By segmenting the model into prediction and modulatory networks, the patent ensures that weight updates in the modulatory network do not affect the prediction network's stable parameters. The prediction network maintains stable representations of old classes while the modulatory network adapts to new classes, resolving the contradiction between learning new knowledge and maintaining stability of old knowledge.
Solution Approach 2:
The modulatory network applies local quality control by generating attention maps that selectively highlight important regions in the input data. This allows targeted updates only in specific weight regions related to new classes, while leaving other regions (old knowledge) unchanged, thus maintaining stability of old knowledge during learning of new knowledge.
3Productivity
If the model uses limited training samples, then the efficiency of learning is improved, but the accuracy of classification deteriorates (overfitting)
Solution Approach 1:
The prediction network performs preliminary action by pre-processing input data and extracting stable features before the modulatory network receives them. This preliminary feature extraction from the prediction network provides a solid foundation that reduces overfitting in the modulatory network when it learns from limited training samples, as the features are already optimized for generalization.
Solution Approach 2:
The attention maps generated by the modulatory network serve as an intermediary mechanism that selectively amplifies informative features from the prediction network's output. This intermediary attention mechanism allows the model to achieve high classification accuracy with limited training samples by focusing computational resources on the most discriminative features, thereby improving learning efficiency without sacrificing accuracy.
Data Source
AI summary
This disclosure provides for methods and system for meta few-shot class incremental learning. According to an aspect a method is provided. The method includes obtaining at least one weight attention map of a first network and updating weights of a second network using the at least one weight attention map, where the second network is a modulatory network. The method further includes generating at least one feature attention map of the second network based on the at least one weight attention map of the first network and a set of input images of at least one class. The method further includes generating at least one feature map of the first network based on the set of input images of the at least one class, and updating the at least one feature map of the first network based on the feature attention map of the second network.