Image may be NSFW.
Clik here to view.
© Dirk Kuiper / ASTRON
Each row shows a different type of FRB morphology, from simple narrowband pulses to scattered, drifting, and complex bursts — including real bursts detected by CHIME/FRB. The first column displays the original bursts, while the remaining columns show reconstructions made by the neural network using increasing numbers of latent variables, or compressed features.
With only one or two latent variables, the network can reconstruct only coarse structures — such as the average intensity and general shape of the burst. But as more latent variables are added (up to ten), the reconstructions progressively improve, recovering finer details like frequency structure, sub-burst drift, and even morphology differences across classes. This illustrates the model’s ability to balance compression and fidelity: capturing key physical features with minimal information.
Remarkably, the IOB-CAE also acts as a denoiser: even bursts with realistic CHIME-like noise are cleaned up in the reconstructions, preserving signal structure while suppressing background noise. This is especially visible in the final row.
These results demonstrate the potential of unsupervised deep learning models for understanding FRB populations. As we move toward a future of hundreds to thousands of bursts per day, tools like the IOB-CAE will help us automatically learn the key features of the data — without needing human labels.
