Philip Graff, NASA
Colloquium
Location
Physics : 401
Date & Time
April 2, 2014, 3:30 pm – 5:00 pm
Description
TITLE: Neural Networks for Machine Learning in Astronomy
ABSTRACT: Artificial neural networks (NNs) are a method of computation loosely based on the structure of a brain. They consist of a group of interconnected nodes, which process information that they receive and then pass this product along to other nodes via weighted connections. A network is able “learn” a relationship between inputs and outputs given a set of training data and can then make predictions of the outputs for new input data. Many calculations in astrophysics are very computationally expensive. If a calculation needs to be performed a large number of times, then it can significantly slow down simulations and analysis. We set out to develop a generic algorithm for the training of NNs in order to facilitate and expedite these repeated difficult computations. I describe the benefits of the SkyNet algorithm and demonstrate its effectiveness in applications to astrophysical problems.
ABSTRACT: Artificial neural networks (NNs) are a method of computation loosely based on the structure of a brain. They consist of a group of interconnected nodes, which process information that they receive and then pass this product along to other nodes via weighted connections. A network is able “learn” a relationship between inputs and outputs given a set of training data and can then make predictions of the outputs for new input data. Many calculations in astrophysics are very computationally expensive. If a calculation needs to be performed a large number of times, then it can significantly slow down simulations and analysis. We set out to develop a generic algorithm for the training of NNs in order to facilitate and expedite these repeated difficult computations. I describe the benefits of the SkyNet algorithm and demonstrate its effectiveness in applications to astrophysical problems.
Tags: