Publications

This paper explores how primordial black holes (PBHs) are potential sources of axions and axion-like particles (ALPs), with a focus on detecting their decay into photons.

• Source of Axions: PBHs emit axions and ALPs as part of Hawking radiation when their temperature surpasses the particle's mass.
• Detection Method: Introduces a new method to track the decay of axions into photons over time on a cosmological scale, predicting the photon spectrum and flux.
• Future Implications: Highlights the potential of upcoming detectors, like e-ASTROGAM, to detect these signals under certain assumptions about PBHs, such as a monochromatic mass spectrum and isotropic distribution.

DeeLeMa is a deep learning network designed for analyzing energy and momentum in high-energy particle collisions, specifically those involving multiple invisible particles.

• Event Analysis: Addresses the challenge in high-energy physics experiments of analyzing collision events with multiple invisible particles.
• Design and Robustness: Built on kinematic constraints and symmetry of event topologies, DeeLeMa effectively estimates mass distribution despite combinatorial uncertainties and detector smearing.
• Flexibility and Impact: Adaptable to various event topologies, leveraging kinematic symmetries, and holds significant potential for advancing high-energy particle collision data analysis.

This study introduces an unsupervised learning approach to assess motion-induced reliability degradation of cardiac volume signals (CVS) in multi-channel electrical impedance-based hemodynamic monitoring.

• Innovative Methodology: Utilizes long-short term memory and variational auto-encoder structures in an encoder-decoder model for self-reproducing CVS input sequences and extrapolating future sequences.
• Advanced Detection and Analysis: Detects low-quality, motion-influenced CVS by comparing input sequences with their neural representations, using a two-sigma rule for determining cutoff values.
• Enhanced Annotation and Error Reduction: Demonstrates that machine-guided annotation can effectively identify motion-induced anomalies with less human error and effort, achieving competitive performance even without labeled data.

This paper introduces novel learning rate schedulers, HyperbolicLR and ExpHyperbolicLR, designed to address the learning curve decoupling problem in deep learning optimization.

• Novel Schedulers: Proposes HyperbolicLR and ExpHyperbolicLR, based on hyperbolic curves, to maintain consistent learning rate changes across varying epoch numbers.
• Performance Consistency: Demonstrates superior stability and consistent performance improvements across increasing epoch numbers in various deep learning tasks.
• Decoupling Mitigation: Introduces the Smoothed Learning Curve Difference metric to quantify and mitigate the learning curve decoupling problem.
• Wide Applicability: Shows versatility across different model architectures and tasks, including image classification, time series prediction, and operator learning.