GIGANTIC: Galactic Interpretive GANs To Identify Curiosities
Advisor(s)
Dr. Joao Caldeira, Fermi National Accelerator Lab
Dr. Brian Nord, Fermi National Accelerator Lab
Location
Room B110
Start Date
26-4-2019 9:45 AM
End Date
26-4-2019 10:00 AM
Abstract
GANs—generative adversarial networks—are neural networks that are composed of two parts, the generative network and the discriminative network, which compete to generate images or other forms of media. In GANs, the generative network produces images, while the discriminative network provides the probability that the image is real. As these networks compete against each other, they are both learning to be much better than normal neural nets, producing higher quality results. Our project is specifically concerned with applying GANs to strong gravitational lensing, a phenomenon occurs when the mass density of the lens is greater than critical density. The gravitational force of the lens can create multiple images, arcs, and Einstein rings. Training GANs on existing data of gravitational lensing could simulate more examples of these images. The other aspect to our project is deblending already existing images of lensing to reduce signal interference and create more accurate pictures. As of now, the work on the networks is still in early stages, but over the course of the next year, we hope to complete functioning nets. When the neural networks are complete, several diagnostics will be run on their efficiency, accuracy, and loss function to determine if they are viable options.
GIGANTIC: Galactic Interpretive GANs To Identify Curiosities
Room B110
GANs—generative adversarial networks—are neural networks that are composed of two parts, the generative network and the discriminative network, which compete to generate images or other forms of media. In GANs, the generative network produces images, while the discriminative network provides the probability that the image is real. As these networks compete against each other, they are both learning to be much better than normal neural nets, producing higher quality results. Our project is specifically concerned with applying GANs to strong gravitational lensing, a phenomenon occurs when the mass density of the lens is greater than critical density. The gravitational force of the lens can create multiple images, arcs, and Einstein rings. Training GANs on existing data of gravitational lensing could simulate more examples of these images. The other aspect to our project is deblending already existing images of lensing to reduce signal interference and create more accurate pictures. As of now, the work on the networks is still in early stages, but over the course of the next year, we hope to complete functioning nets. When the neural networks are complete, several diagnostics will be run on their efficiency, accuracy, and loss function to determine if they are viable options.