Jonah Goldfine, MS

About

I am an astronomer completing a Masters of Research studying the use of machine and deep learning to improve the capabilities of astronomical instruments. I currently live in Sydney and study at Macquarie University and previously studied at New York University and the University of Denver. I grew up Scottsdale, Arizona with my family — including my dogs.

Despite growing up in a family of current and future lawyers, I was always interested in science. Whether it was through sci-fi like Star Wars, shows like Bill Nye the Science Guy or Mythbusters, and actually being able to pay attention during science courses, I was destined for STEM, but it was a summer camp in northern Wisconsin where I really first fell in love with the stars. This interest grew in high school while taking an Astronomy course. The summer camp is also where my interest in outreach began.

Beyond astronomy, I love sci-fi and fantasy media and am always accepting book and movie recommendations to add to my ever-expanding list of things to read or watch. I like to spend my time playing board games and Dungeons & Dragons with friends. I am also a massive aquariums and zoos. FInally, I love to bake. I really enjoy figuring out the science of baking, developing my own recipes, and sharing what I make.

Pets and Animal Friends

Research

Current Work

For my Masters of Research, I am working with Prof. Benjamin Pope (Macquarie University) and Ph.D. candidate Max Charles (University of Sydney) to use machine learning to augment dorito, a regularized-maximum likelihood framework for deconvolving interferograms from JWST's Aperture Mask Interferometer (AMI) with the goal of improving the angular resolution in order to see more complex structures within protoplanetary disks. In particular, we hope to find complex substractors that could indicate the presence of planets and potential dynamical processes that form them.

In order to do this, I have developed a convolutional autoencoder that learns two things:

• How to encode the two dimensional image space into a compressed representation, or latent space, and then decode it back into the original image space.
• Denoise images as it is learning the encoding and decoding.

By making the encoder and decoder portions of the autoencoder seperable, I am able to replace the pixel space portion of the deconvolution within dorito with one in the machine-learned latent space. This means that there are far fewer features to gradient descend on. Like many other data compression methods, this is not lossless, but a vast majority of the information is preserved. Going one step further, I've used a normalized version of the latent space to create a machine-learned initialization for dorito to replace something like a dirty beam or uniform distribution.

So far I've developed an autoencoder trained on emnist leters and numbers as a baseline architecture before switching to images that better represent the kinds of objects we might be able to see with AMI.

Past Work

During my time at NYU and the University of Denver, I worked with Professor David W. Hogg (NYU, Flatiron Institute) and Adrian Price-Whelan (Flatiron Institue) looking for a correlation between light-curve variability and Monte Carlo samples using The Joker (Yo-ker) and the Lomb-Scargle and box- least-squares periodograms. This resulted in two abstracts and posters presented at the 241st and 243rd AAS meetings.

This project involved taking periods of variable stars using various method (The Joker and periodograms) to determind the optimal period to "fold" the lightcurve. I would also compare the periodgram-folded lightcurves with the period-eccentricity samples from The Joker.

Timeline & CV

Academic Timeline

Research Timeline

Teaching Timeline

CV

• Download CV

Contact

• Bluesky
• Github
• LinkedIn
• ORCiD
• MQ Email

Elements

i = 0;

while (!deck.isInOrder()) {
    print 'Iteration ' + i;
    deck.shuffle();
    i++;
}

print 'It took ' + i + ' iterations to sort the deck.';