Searching the Cosmos with Neural Networks

Motivation

• The amount of data that telescopes produce is growing…astronomically
  • LSST to collect 20 TB per night starting in 2022
  • Square Kilometre Array will generate 2 PB daily starting in 2028
• One of the largest sky surveys currently, SDSS, has collected 40 TB over the past 20 years
• New systems and methods are necessary to process and keep up

Gravitational Lenses

Gravitational Lenses

Convolutions

Feature Extraction

• Different kernels can highlight different features in an image

Convolutional Neural Networks

• The networks “learns” the best kernel weights
• Generally use many kernels and combinations of kernels

Challenges

• Choosing the convolution model
  • How many kernels to use each step of the way?
  • What size should they be?
  • How do decide when to pool or stride?
  • Lots of practice to get a feel for what seems to work well and what does not.
• Acquiring and reading in training data
  • Need labeled data for a supervised algorithm like this
  • Often times large datasets, so can only read in a bit at a time
  • Practice working, cleaning, and streaming in large datasets
• Dealing with poor fitting or overfitting
  • Practice looking at different analytics to determine how your model is performing

Interested in More?

• Talk to me!
  • My office is Collins 311
  • Email: jjrembold@willamette.edu