UCLA SETI Group newsletter 2021-10-25

View this email in your browser

October 25, 2021

Dear friends,

I am thrilled to report that UCLA physics graduate student Paul Pinchuk successfully defended his PhD dissertation this summer. He became one of only a dozen earthlings to obtain a PhD in the field of SETI. Paul's contributions have propelled the field in a number of ways. With an algorithm based on the concept of topographic prominence – the height of a peak above the lowest contour line encircling it – Paul multiplied the number of narrowband detections in radio data by a factor of approximately 10. To my knowledge, his signal injection and recovery analysis provided the first estimate of the detection efficiency of a SETI data processing pipeline, an estimate that is key to any calculation purporting to provide an upper bound on the prevalence of extraterrestrial radio transmitters. His machine learning application based on a convolutional neural network (Figure 1) reduced the visual confirmation burden of promising signals by a factor of 10. Paul's assistance with the UCLA SETI course has also been terrific, as observed from my vantage point and confirmed by student evaluations. Throughout the years, Paul exhibited an entrepreneurial spirit and expressed a preference for a career in industry. I look forward to hearing about Dr. Pinchuk's future accomplishments.

Figure 1: Architecture of the convolutional neural network designed and tuned by Dr. Paul Pinchuk for the UCLA SETI Group based on the Xception architecture.

The ability to train a neural network to accomplish a specific task is transformative. In our case, we provided a million pairs of images to a neural network with corresponding true or false labels. Each label gives the answer to a simple question: is the signal observed in the first image also present in the second image? Although it may seem trivial, the answer to this question provides one of the most powerful ways to distinguish a technosignature from radio frequency interference (RFI). You may recall that our observations are structured with a specific scan sequence: source A – source B – source A – source B. When we analyze the data, we verify whether signals observed in the direction of source A are also observed in the direction of source B. A signal that is observed in both directions is classified as RFI and removed from consideration because a genuine technosignature would originate from only one direction on the sky. The primary beam of the telescope is sufficiently narrow that it does not admit signals from other directions. This "direction-of-origin" filter generally relies on estimates of signal properties such as frequency and frequency drift rate to match corresponding signals. Our existing implementation is >99.5% efficient, but with millions of detections per hour of telescope time, the remaining 0.5% leaves thousands of signals in need of visual confirmation. Paul thought that a neural network that analyzes images directly could supplement our existing filter that analyzes estimates of signal properties, and indeed his machine learning application improves the overall classification efficiency to 99.95%. The neural network was trained (i.e., its numerous coefficients were optimized) by analyzing our labeled training set. When the network processes new data, it provides answers with precision and recall values of 99.15% and 97.81%, respectively. Importantly, the classification is fast, suggesting that near-real-time processing is possible.

Table 1: Summary of the UCLA SETI Group's observations between 2016 and 2021. For context, a recent survey by the Breakthrough Listen project with the same telescope, receiver, and detection threshold had 1327 primary targets, 37.1 million detected signals, and a hit rate density of 0.11 sig kHz^-1 hr^-1(Price et al. 2020).

It is rewarding to think about how far we have come since we launched the UCLA SETI Group in 2016 with Janet Marott's initial gift. We have sampled over 36,000 stars with known distances and detected over 56 million candidate signals (Table 1). We have developed algorithms that perform 100+ times better than those of other pipelines. We have implemented a powerful machine learning classifier and a signal injection tool that is necessary to place meaningful upper limits on the prevalence of extraterrestrial transmitters. And we have exposed more than 100 undergraduate students and more than 10 graduate students to SETI science in a university course that appears to be the first of its kind. We haven't found evidence of life elsewhere yet, but we are constantly refining and expanding our search. Our mission would not be possible without your continued generous support, for which I am extremely grateful.

Warm regards,

Jean-Luc Margot