Early-fusion Based Pulsar Identification with Smart Under-sampling

TL;DR

This paper introduces a novel deep learning framework for pulsar identification that combines smart under-sampling, improved loss functions, and early-fusion of features, significantly enhancing speed and accuracy in handling imbalanced data.

Contribution

The study is the first to integrate smart under-sampling, an improved loss function, and early-fusion strategies in pulsar recognition, achieving faster training and competitive accuracy.

Findings

Training time reduced by 10X compared to previous methods.

Achieved high F1 score indicating effective pulsar recognition.

Outperformed existing methods in both speed and accuracy.

Abstract

The discovery of pulsars is of great significance in the field of physics and astronomy. As the astronomical equipment produces a large amount of pulsar data, an algorithm for automatically identifying pulsars becomes urgent. We propose a deep learning framework for pulsar recognition. In response to the extreme imbalance between positive and negative examples and the hard negative sample issue presented in the HTRU Medlat Training Data,there are two coping strategies in our framework: the smart under-sampling and the improved loss function. We also apply the early-fusion strategy to integrate features obtained from different attributes before classification to improve the performance. To our best knowledge,this is the first study that integrates these strategies and techniques together in pulsar recognition. The experiment results show that our framework outperforms previous works with the respect to either the training time or F1 score. We can not only speed up the training time by 10X compared with the state-of-the-art work, but also get a competitive result in terms of F1 score.