Please use this identifier to cite or link to this item:
Title: A weight-normalization based causal classifier for long-tailed classification
Authors: An, Zheyuan
Keywords: Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
Issue Date: 2021
Publisher: Nanyang Technological University
Source: An, Z. (2021). A weight-normalization based causal classifier for long-tailed classification. Final Year Project (FYP), Nanyang Technological University, Singapore.
Project: SCSE20-0376
Abstract: The success of deep learning techniques in computer vision is largely supported by the availability of large-scale datasets. However, it is difficult to maintain a balanced dataset as the dataset size grows because a few head classes will appear much more frequently compared to a large number of tail classes. Therefore, long-tailed classification is crucial to large scale computer vision tasks. However, many common solutions to long-tailed classification rely on changing the original distribution of the classes in the dataset, causing the information of class structures to be lost. In this paper, we propose a method using a weight-normalization based causal classifier to tackle the long-tailed classification under a causal framework. Specifically, our approach disentangles the magnitude and direction of the weight vectors of the classifier to allow for causal intervention on each of their effects. The model was trained and tested on the Long-tailed CIFAR-10/100 datasets and was able to outperform previous approaches in highly imbalanced datasets.
Schools: School of Computer Science and Engineering 
Fulltext Permission: restricted
Fulltext Availability: With Fulltext
Appears in Collections:SCSE Student Reports (FYP/IA/PA/PI)

Files in This Item:
File Description SizeFormat 
FYP_Report_SCSE20-0376_An Zheyuan.pdf
  Restricted Access
446.46 kBAdobe PDFView/Open

Page view(s)

Updated on Jun 8, 2023


Updated on Jun 8, 2023

Google ScholarTM


Items in DR-NTU are protected by copyright, with all rights reserved, unless otherwise indicated.