Research Area:  Machine Learning

Abstract:

Few-shot learning aims to discriminate images from novel categories using only a few available training examples. While existing few-shot methods can adapt quickly and precisely to new categories, they often struggle to retain knowledge of base categories that were used in the training phase. To address this challenge and support lifelong learning, generalized few-shot learning has been introduced to enable few-shot models to classify both base and novel categories. However, as the number of categories increases, few-shot models can lose efficiency due to the limited amount of visual information available for each category. To address this limitation, we propose the knowledge-augmented weight generation (KAWG) method, which incorporates semantic information in addition to visual features. Specifically, KAWG combines textual descriptions and entity relationships extracted from knowledge graphs and visual features to generate more robust classifiers for generalized few-shot learning tasks. Through our meta-training strategy, KAWG can retain the knowledge learned from base categories to the greatest extent when transferring to novel classes. Experiments show that our approach achieves state-of-the-art performance on some generalized few-shot benchmarks.

Keywords:  

Author(s) Name:  Dianqi Liu, Liang Bail, Tianyuan Yu

Publisher name:  Springer

DOI:  10.1007/s11063-023-11278-1

Volume Information:  Volume 55, pages 7649-7666, (2023)

Paper Link:   https://link.springer.com/article/10.1007/s11063-023-11278-1#:~:text=Generalized%20few%2Dshot%20learning%20aims,when%20predicting%20X_T%20%5Crightarrow%20Y_T.