Few-Shot Learning Method Using Multi-Task Neural Network
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current few-shot learning methods face challenges with overfitting and performance deterioration when classifying complex data, especially in environments with limited labeled data, and require separate labeled datasets for separate learning and data selection.
Innovation Solution
A method that involves multiple learning tasks: meta-training with a first model, supervised learning with a second model sharing the main body, weakly-supervised learning with a third model generating pseudo labels, and unsupervised learning with a clone model, all using an artificial neural network structure to improve feature representation and reduce the need for extensive labeling.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If typical supervised learning is applied with limited labeled data, then the learning process can be performed, but overfitting occurs and model performance deteriorates
Solution Approach 1:
The patent applies preliminary action by performing meta-learning before actual few-shot learning tasks. The model is pre-trained on meta-training data with multiple tasks to learn generalizable features and adaptation strategies, so that when faced with limited labeled data in actual applications, the model already possesses robust feature representations that prevent overfitting
Solution Approach 2:
The patent changes parameters by introducing multiple learning rates for different model components (main body vs. head), applying different regularization strengths, and adjusting the number of training epochs for meta-learning versus task-specific learning. These parameter changes enable effective learning with limited data while preventing overfitting
2Reliability
If separate learning is performed with a separate labeled dataset, then model performance can be improved, but the complexity of data preparation and selection increases
Solution Approach 1:
The patent merges the meta-learning process and task-specific learning into a unified framework. The same model architecture (main body and head) is used for both meta-training and actual few-shot tasks, eliminating the need for separate learning systems and reducing data preparation complexity while maintaining performance improvements
Solution Approach 2:
The patent creates a universal model that performs multiple functions: it can learn from meta-training data across different tasks and then adapt to new few-shot tasks with limited data. This multi-functional model eliminates the need for separate specialized learning systems for different data scenarios
Data Source
AI summary
A few-shot learning method according to an embodiment may be performed in a computing device including one or more processors and a memory storing one or more programs executed by the one or more processors. The method may include performing a first task for subjecting a first model to few-shot learning (FSL) based on one or more meta-training data and performing a second task for subjecting a second model to supervised learning based on one or more derived data modified from the one or more meta-training data.