Ensemble Deep Learning for Real-Bogus Classification with Sky Survey Images.

Area of Science:

• Astronomy and Astrophysics
• Machine Learning
• Bio-inspired Computing

Background:

• The simultaneous detection of gravitational waves (GWs) and electromagnetic counterparts, like GW170817, highlights the importance of observing transient astronomical events.
• Neutron star mergers emit multi-frequency waves, but rapid localization of these events for follow-up studies remains challenging.
• Traditional methods struggle to efficiently identify short-lived transient events within massive sky survey datasets, such as those from the Gravitational-Wave Optical Transient Observer (GOTO) project.

Purpose of the Study:

• To develop an advanced computational methodology for enhancing the classification of astronomical transient events.
• To leverage deep learning, specifically Convolutional Neural Networks (CNNs), inspired by biological vision systems, for improved transient detection.
• To create a scalable and robust system for real-time analysis of transient phenomena in large-scale astronomical surveys.

Main Methods:

• Implementation of Convolutional Neural Networks (CNNs) with a bio-inspired architecture mimicking hierarchical visual processing in animal brains.
• Utilization of transfer learning and fine-tuning on ImageNet models for adaptive learning with limited astronomical data.
• Application of data augmentation techniques (rotation, flipping, noise injection), regularization (dropout), and ensemble learning (Soft Voting, Weighted Voting) to improve model generalization and robustness.

Main Results:

• The proposed bio-inspired deep learning framework significantly enhances the precision and reliability of astronomical transient detection.
• The methodology demonstrates effective automatic identification of complex spatial patterns in astronomical images.
• The system provides a scalable solution suitable for real-time processing of data from extensive sky surveys like GOTO.

Conclusions:

• Deep learning, particularly CNNs inspired by biological systems, offers a powerful approach to overcoming challenges in transient event detection.
• This study validates the effectiveness of bio-inspired computational strategies in astrophysics for real-time analysis and discovery.
• The developed framework promises to advance the study of transient astronomical phenomena, enabling faster and more accurate follow-up observations.