Detection of cosmic objects and astronomical events is only a half of the job. The other half is probably no less intricate: how to recognize interesting things among the ‘piles’ of noisy data? It appears, that astronomical observations are not so much different from other experiments involving large volumes of information. For example, it is even possible to use automated data processing and pattern recognition methods to detect stars, galaxies, black holes and other fascinating spots of the Universe. Well, at least it works for pulsars now.
A large team of scientists involving specialists from United States, Canada, Germany, Australia, United Kingdom and Netherlands proposed a neural network-based software system which is capable of identifying pulsars in noisy data obtained from various astronomical surveys. Authors of this development also published a paper on arXiv.org, which presents their Pulsar Image-based Classification System (PICS) in more detail.
According to researchers, PICS operates using principles of artificial intelligence (AI). Certainly, this so-called artificial intelligence isn’t the one we could imagine: it is based on the artificial neural networks, which in this case are used to process stellar signal data from candidate locations observed by terrestrial or space-based observatories.
As any other human-developed neural network, the PICS system needs training in order to accomplish the machine-learning procedure. In this case the samples of known data, including the previously spotted pulsars from Pulsar Arecibo L-band Feed Array Survey, were fed into the automated classification program prior to performing a new analysis.
Compared to the similar existing systems, the PICS is also able to detect specific features common to pulsars of most frequent types during the learning phase. The other tools for automated pulsar detection need a separate procedure and pre-designed data patterns just to configure them.
The PICS performance verification experiments were carried out using data from the other pulsar survey, the Green Bank North Celestial Cap survey. The algorithm recognized correctly 264 out of 277 pulsar-like objects, including all 56 previously known pulsars, hereby achieving the overall identification precision of 95.3%.
According to the authors, pulsar recognition rate of the Pulsar Image-based Classification System should increase over time, when larger data sets will be collected and used to train the neural network used for this purpose. Thus we can expect, that in the nearest future the astronomical survey databases will be filled with lots of newly detected pulsars. And perhaps a similar system will be developed to detect other distinct objects of our Universe.
By Alius Noreika, Source: Technology.org