Balanced Distributed Augmentation for Multi-Label Few Shot Learning with Prototypical Network

Many methods have been presented as a few shot learners in order to enhance few-shot learners. Some of these methods involve routine-based pre-trained language models and novel pipeline for automating the prompt generation. In this study, we propose a new evenly distributed data augmentation techniq...

Full description

Saved in:
Bibliographic Details
Published in2022 30th Signal Processing and Communications Applications Conference (SIU) pp. 1 - 4
Main Authors Haruna Mohammed, Hamza, Oner, Alper
Format Conference Proceeding
LanguageEnglish
Published IEEE 15.05.2022
Subjects
Distributed databases
few shot learning
Labeling
meta-learning
multi-label classification
Pipelines
Probabilistic logic
prototypical network
Semantics
Sentiment analysis
Signal processing
text data augmentation
Online AccessGet full text

Cover

More Information
Summary:Many methods have been presented as a few shot learners in order to enhance few-shot learners. Some of these methods involve routine-based pre-trained language models and novel pipeline for automating the prompt generation. In this study, we propose a new evenly distributed data augmentation technique, which generates samples according to the probabilistic distribution of the relationship of each label with the mean of a label group. In the labeling phase, we present a semantic sentiment analysis approach in order to increase the realism of the data, in a more semantic augmentation way. The results show that this approach improves the few shot learners. In addition to this, we compare our adaptation approach to other traditional problem transformation methods. The newly developed approach outperforms these traditional methods, especially when the classifier learns from a limited number of samples.
DOI:10.1109/SIU55565.2022.9864875