Major
Physics
Research Abstract
Measuring gravitational lensing by galaxies is the only way to directly study the elusive dark matter. However, gravitational lensing is a very rare phenomenon (~1 in 10,000 galaxies). Our goal is to find new strong gravitational lenses using deep neural networks (“neural nets”). We train our neural nets on a hand-labeled set of images, consisting of both lenses and non-lenses (“the training sample”). We then apply the trained neural nets to a “validation set” to assess the accuracy and precision of its predictions. Given the rarity of lenses, we cannot tolerate a false positive rate higher than 0.1%. This is to minimize or possibly eliminate human inspection. This is an extremely high bar for Machine Learning (“ML”) algorithms. Our data sets are selected from real observational data. Utilizing real data has not been attempted before. In this project we update and modify an existing neural net model, originally created by a team at Carnegie Mellon University (CMU) written in Theano, a python library used for ML. After training on real, observed data the neural network recommended ~40,000 recommendations from a sample of 15 million galaxies. All recommendations were inspected by hand, from which there were hundreds of high probability candidates for strong lensing.
Faculty Mentor/Advisor
Xiaosheng Huang
Included in
Finding Strong Gravitational Lenses with Residual Neural Networks
Measuring gravitational lensing by galaxies is the only way to directly study the elusive dark matter. However, gravitational lensing is a very rare phenomenon (~1 in 10,000 galaxies). Our goal is to find new strong gravitational lenses using deep neural networks (“neural nets”). We train our neural nets on a hand-labeled set of images, consisting of both lenses and non-lenses (“the training sample”). We then apply the trained neural nets to a “validation set” to assess the accuracy and precision of its predictions. Given the rarity of lenses, we cannot tolerate a false positive rate higher than 0.1%. This is to minimize or possibly eliminate human inspection. This is an extremely high bar for Machine Learning (“ML”) algorithms. Our data sets are selected from real observational data. Utilizing real data has not been attempted before. In this project we update and modify an existing neural net model, originally created by a team at Carnegie Mellon University (CMU) written in Theano, a python library used for ML. After training on real, observed data the neural network recommended ~40,000 recommendations from a sample of 15 million galaxies. All recommendations were inspected by hand, from which there were hundreds of high probability candidates for strong lensing.
