Fast identification of strong gravitational lenses with deep learning on FPGAs and other heterogeneous computing devices
Abstract
We present our work in pursuit of real-time classification of strong gravitational lenses in future wide-field cosmic surveys, like LSST and WFIRST. Strong lenses are maturing as cosmic probes of large-scale structure and cosmic expansion, and future surveys have promise in realizing their potential. In this work, we employ deep neural networks to classify simulated gravitational lensing data and train our networks on 20,000 images of lenses and non-lenses with 4 bands of data. We feed these images into state-of-the-art Convolutional Neural Networks (CNNs), like ResNet50. We then perform inference on 100,000 images on various hardware (e.g. CPUs, GPUs, TPUs, and FPGAs) and examine the overall accuracy and efficiency of each model architecture and hardware combination. The aim is to develop a proof-of-concept model-architecture-hardware pipeline to perform fast inference at or near LSST when the telescope begins collecting data in the next few years. Our initial ResNet50 model achieves an accuracy of ~75% on a GPU (with a total runtime of ~4 hours) using 60,000 images and minimal pre-processing, but by increasing the size of our dataset as well as performing more aggressive pre-processing techniques on the images, we believe our accuracy will improve significantly. We will show the latest results for other models and will use the best combination of algorithms and hardware to construct the ideal pipeline for analyzing wide-field survey images. We expect that our automatic classification pipeline will be efficient enough (likely on the order of milliseconds/image) that fast, just-in-time follow-up observations can be performed on images our program denotes as containing gravitational lensing. Fast detection of lenses is critical for allowing the scientific community to capture more lensing data (particularly of lensed quasars and supernovae), which will improve our understanding of dark energy and spacetime expansion.
- Publication:
-
American Astronomical Society Meeting Abstracts #235
- Pub Date:
- January 2020
- Bibcode:
- 2020AAS...23530303C