Semantic-guided distribution calibrating for few-shot classification
The meta-learning paradigm is a mainstream method to solve the problem of few-shot learning. It constructs multiple tasks through episodic training so that the model can quickly generalize to new tasks. However, episodic training only sample a few training instances (shots) per class for simulating...
Saved in:
Main Author: | |
---|---|
Other Authors: | |
Format: | Thesis-Master by Coursework |
Language: | English |
Published: |
Nanyang Technological University
2022
|
Subjects: | |
Online Access: | https://hdl.handle.net/10356/155041 |
Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Institution: | Nanyang Technological University |
Language: | English |
Summary: | The meta-learning paradigm is a mainstream method to solve the problem of few-shot learning. It constructs multiple tasks through episodic training so that the model can quickly generalize to new tasks. However, episodic training only sample a few training instances (shots) per class for simulating the test-
time scenario, the randomness of sampling and scarcity of data will bring biased feature distribution of the task, which occurs in both meta-training and meta-testing. In this work, we propose a semantic guided adaptive distribution
calibrated framework to alleviate this phenomenon, committing to using debiased information to eliminate bias. Semantic Conditional Batch Normalization (SCBN) , is to alleviate the deviations of the same class distribution in different episodic caused by sampling from the level of feature maps. In addition, a semantic Transformer (ST) is introduced to map the support and query set into a new feature space, where their domain gap is smaller. In both modules, the category semantic representation (text or language) is used as a high-level feature to modify the distribution. In this way, debiased semantic representation can make the model not be disturbed by sampling, and adjust the feature distribution adaptively, so that the class distribution can be separated as much as possible in the feature space. Experimental results show that it compares favourably to the previous state-of-the-art results for images only and images combined with semantics-based approaches. |
---|