Active Galactic Nuclei Multi-Wavelength X-ray Spectral Analysis Using Neural Networks
Presenter: Professor Shiqi Yu (Research Assistant Professor, Department of Physics and Astronomy, University of Utah)
Abstract:
Accurate parameter estimation in X-ray astronomy typically relies on traditional methods, such as likelihood-based spectral fitting, which can be computationally prohibitive as model complexity and data dimensionality increase. In this talk, I present a neural network-based framework designed to bypass iterative fitting by directly mapping spectral observations to physical parameters. Using the Circinus galaxy as a benchmark, we demonstrate how architectures trained on synthetic data from theoretical models can recover intrinsic properties, such as column density and torus geometry, with both high speed and high precision. This approach maintains the physical rigor required for broadband analysis with multiple telescopes while significantly reducing inference time. I will discuss the challenges and resolutions associated with training and predicting on multi-instrument data, as well as the potential for these AI-driven methods to enable large-scale systematic studies across various astrophysical sources and other scientific applications.
Bio: Prof. Yu is a Research Assistant Professor in the Department of Physics and Astronomy at the University of Utah. Her research is centered on multi-messenger astrophysics, where she leverages machine learning as a powerful tool to investigate high-energy neutrino sources and complex astrophysical phenomena. As a member of the IceCube Collaboration and former co-lead of the Reconstruction and Machine Learning working group, she has extensive experience applying machine learning techniques to diverse astrophysical data. This work aims to enable high-speed and high-precision inference using multi wavelength data to further our understanding of the high energy universe.
Controlling LLM's via Activation Geometry
Presenter: Amirali Abdullah (Lead AI Researcher, Thoughtworks Inc)
Abstract: Controlling the behavior of large language models at inference time is an increasingly important problem. In this talk, I present a simple and unified approach to steering model behavior based on activation geometry. By learning a single classifier over hidden representations, we can derive directions that control multiple attributes such as helpfulness, style, or safety, and compose them dynamically without retraining.
This framework enables flexible, low cost control of model outputs and highlights a geometric view of representation space beyond fixed linear directions. I will discuss empirical results showing how this approach supports multi attribute control in practice, and briefly outline how such steering mechanisms can be useful in scientific settings where reliable and interpretable model behavior is critical. Our recent followup work suggests that similar activation level interventions can extend across modalities, enabling systematic analysis and control in text to image models through composable operations.
Bio: Amirali Abdullah is a Lead AI Researcher at Thoughtworks Inc and a Research Advisor at Martian Learning. His research focuses on the interpretability and control of large language models, with particular emphasis on activation-level steering, representation geometry, and the structure of learned features.