Synthetic Labeling: A Novel Approach to Advancing Few-Shot Learning

Zhaoyan Lyu, Gholamali Aminian, Miguel R. D. Rodrigues

Keywords: Few-shot learning, regularizer, classification, synthetic labeling

TL;DR: This paper introduces an innovative synthetic labeling algorithm to enhance generalization performance in classification tasks.

Abstract: In the field of few-shot learning, the scarcity of labeled data significantly hinders progress. This paper introduces an innovative regularization algorithm designed to enhance generalization performance in classification tasks by leveraging synthetically labeled data. The approach utilizes a single encoder and multiple decoders, trained on both an original dataset with ground truth labels and synthetic datasets with artificial labels. Our empirical studies demonstrate that this method effectively improves neural network generalization, both independently and when integrated with other regularizers. This versatility underscores the potential of synthetic labeling in overcoming data limitations in few-shot learning scenarios.

Submission Number: 140