Can we teach a machine to recognize what we can’t? This may sound like an obvious yes to astronomers, because astronomers are constantly working in regimes where the signal is the same order of magnitude as the noise, and we often need to manipulate the data to extract a measurement that was not apparent with our own eyes. To someone working in machine-learning, where to goal is often to teach (or train) a machine to perform simple recognition tasks that human do effortlessly, the answer to this question would probably be not without a lot of data, and they’d also be right. What happens, however, when we don’t have enough examples of a type of object to train on, or even scarier, when the experts have a difficult time recognizing the objects even when they know they are there?
My colleagues and I are facing this problem right now. To discuss the problem, let me give a little background.
The newly constructed Australia Square Kilometre Array Pathfinder Telescope (ASKAP) will be one of the most powerful survey radio telescopes in the work, operating between 700-1800 MHz in both continuum (total brightness) and spectral-line (brightness as a function of frequency) modes. The telescope will conduct a dedicated polarization survey (called POSSUM), with the primary goal of detecting Faraday rotation measures (RM) from background radio sources. The RM, which measures the integrated amount of thermal electrons and magnetic field along the line of sight, can hold the key to unraveling several mysteries of cosmic magnetism. One problem that we have in preparing a catalog of RMs for polarized sources, is distinguishing between those that are a simple “Faraday thin” source or a more complex source with multiple components or “Faraday thick”. Below I show an example of a simple Faraday thin source, where the polarization angle of the (radio) light has been rotated by a simple cloud of material.
The problem is, it has been shown that polarized sources with two, closely spaced Faraday thin components (and are thus complex) can look like a Faraday thin source, and further, the RM that is measured will be different than the two individual RMs (and it might not be the average of them either). I’ll describe next how I think we can approach this problem with deep convolutional neural networks.