This is *really* impressive. The Kepler Telescope has identified 2525 confirmed exoplanets - many of which are Earth-sized. The total confirmed number of exoplanets is 3567. What is particularly interesting, is that the machine-learning
tool used was a TensorFlow application (very similar to MINST-type classification), which used a training set of roughly 15,000-case observations. This is a sufficiently small enough training set that the entire model can run on a desktop personal computer
- total training time was only a few hours.
What this shows is that machine-learning can be used to effectively to augment and extend the ability of human astronomers in a elegant and effective way, as the Kepler dataset is already large and will
continue to grow. Given we have training cases where humans accurately found planets, we can apply the machine-learning tool to expand the search, and specifically focus on weak signals which are typically the kinds of signals that smaller, Earth-size
planets will create.
The researcher are: Chris Shallue & Andrew Vanderburg. Their paper on this research is at: https://www.cfa.harvard.edu/~avanderb/kepler90i.pdf
They are using logistic neurons, I notice. Their TensorFlow
model will be released to the open-source community, according to the NASA press conference. Looks like very good work, which can find immediate application in other areas.