Meta-Learned Semantic Mapping for Incremental Few-Shot Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing incremental few-shot learning methods face challenges in learning novel classes without available semantic information, as they require semantic information as correct answer data, limiting their applicability.
Innovation Solution
A machine learning device and method that includes a semantic prediction unit trained in a meta-learning process to predict semantic information from input images, allowing learning of novel classes with or without added semantic information, by generating pseudo few-shot classes and using the trained unit's predictions as correct answers.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If semantic information is required as correct answer data for learning novel classes, then learning precision can be maintained, but the method cannot learn images without semantic information
Solution Approach 1:
The system enables the model to generate its own correct answer data by predicting semantic information from base class images. The semantic prediction unit uses learned semantic patterns to autonomously create supervision signals for novel class learning, eliminating the need for external semantic annotations.
Solution Approach 2:
The system performs preliminary learning of semantic information patterns during base class training. By pre-learning the relationship between images and their semantic characteristics on base classes, the model prepares the semantic prediction unit to generate accurate predictions for novel classes without requiring their semantic annotations.
2Reliability
If new tasks are learned using another small amount of parameters, then catastrophic forgetting is avoided, but learning efficiency with limited data is constrained
Solution Approach 1:
The semantic prediction unit acts as an intermediary that bridges base class knowledge and novel class learning. It translates learned semantic patterns into predictions that serve as correct answer data, enabling efficient few-shot learning while preserving base class knowledge through the semantic abstraction layer.
Solution Approach 2:
The system changes parameters by using semantic predictions as dynamic correct answer data instead of fixed annotated semantics. This parameter transformation enables the model to adapt to novel classes with minimal data while maintaining stability through the semantic representation learned from base classes.
Data Source
Figure 1
Figure 2A~2C
Figure 3
AI summary
A feature extraction unit (220) extracts a feature vector from input data. A semantic prediction unit (250) is a module has been trained in advance in a meta-learning process and that generates a semantic vector from the feature vector of the input data. A mapping unit (230) is a module that has learned a base class and that generates a semantic vector from the feature vector of the input data. An optimization unit (240) optimizes parameters of the mapping unit using the semantic vector generated by the semantic prediction unit as a correct answer semantic vector such that a distance between the semantic vector generated by the mapping unit and the correct answer semantic vector is minimized when semantic information is not added to input data of a novel class at the time of learning the novel class.