CSE Seminar Announcement

Deep Learning with HPC Simulations for Time-Series Signal Processing Applied to Gravitational Wave Astrophysics

Wednesday, October 25, 2017
1030 National Center for Supercomputing Applications (NCSA)
12:00pm - 1:00pm
Lunch Provided

Daniel George
CSE Fellow
Astronomy

Abstract

A new era of gravitational wave (GW) astronomy has begun with the recent detections by LIGO. However, current data analysis pipelines are limited by the extreme computational costs of template-matching, thus facing significant delays and inability to detect all GW sources. I will start with an introduction to deep learning with artificial neural networks. I will then describe the highly scalable Deep Filtering technique, based on two 1D convolutional neural networks, that I developed recently to resolve these issues, which allows real-time detection and parameter estimation of GW signals whose amplitudes are much weaker than the background noise. This initiates a new research paradigm using data derived from high-performance physics simulations on supercomputers, including Blue Waters, to train innovative artificial intelligence algorithms that exploit emerging hardware architectures such as deep-learning-optimized GPUs. I will also discuss my recent work on applying transfer learning and unsupervised clustering methods for classifying anomalous noise transients in spectrogram images of LIGO data. I will conclude by discussing my ongoing research and future plans including new deep learning methods for denoising LIGO data with recurrent neural network auto-encoders and generative modeling of GW signals. This deep learning framework for low-latency analysis of the raw big data collected by observational instruments can enable real-time multimessenger astrophysics, which promises groundbreaking insights about the universe.