Machine learning in Nebula

Message boards : Nebula : Machine learning in Nebula
Message board moderation

To post messages, you must log in.

AuthorMessage
Profile David Anderson
Volunteer moderator
Project administrator
Project developer
Project scientist
Avatar

Send message
Joined: 13 Feb 99
Posts: 173
Credit: 502,653
RAC: 0
Message 1978639 - Posted: 4 Feb 2019, 21:52:42 UTC
Last modified: 4 Feb 2019, 21:55:01 UTC

There's considerable interest in using ML techniques, such as neural networks and cluster analysis, in SETI. Can we use ML in SETI@home, and in particular in Nebula? Let's look at the two main functions of Nebula:

[Caveat: I'm not an ML expert. If you are, and have ideas or suggestions, let me know].

1) RFI removal.
This involves going through the ~10 billion signals and identifying sets of signals that, taken together, have properties typical of terrestrial and not celestial origin (e.g. closeness in time and frequency but not position). We've developed several algorithms that do this, and they seem to work pretty well. It's possible that ML could do better, but so far we haven't figured out how to use it for this purpose. The basic problem is lack of training data. If we trained a network based on the outputs of our current algorithm, it wouldn't "learn" anything beyond what the algorithms already do.

However, there are probably approaches we haven't thought of.

2) Multiplet finding and scoring.
There are two related parts here. A "scoring function" takes a set of signals and estimates the probability that it's from ET - i.e. the probability that it's not noise. It does this by computing a number of properties of the set - e.g. its closeness in position, its power, and so on - and combining these with different weights.

The "multiplet finding algorithm" goes through the ~10 billion signals, and scans all sky positions and frequency bands.
In each position/frequency neighborhood it outputs a subset of signals that's likely to produce a high multiplet score.

In the case of multiplets, we do have a potential source of training data, namely birdies (simulated ET signals). We want the system to be good at finding birdies: i.e. finding multiplets made up mostly of birdie signals and with high scores relative to other multiplets.

One possible application of ML is to make an intelligent scoring function, based on a neural network. The inputs to this network would be the various properties of a multiplet, and its output would be a score. It would be trained using a mixture of birdie and non-birdie multiplets.

Of course, the multiplet finding algorithm assumes the current scoring function. We'd want to look at how changes in the scoring function should feed back into the finding algorithm
ID: 1978639 · Report as offensive
Profile Byron Leigh Hatch @ team Carl Sagan
Volunteer tester
Avatar

Send message
Joined: 5 Jul 99
Posts: 4548
Credit: 35,667,570
RAC: 4
Canada
Message 1978730 - Posted: 5 Feb 2019, 8:20:40 UTC - in response to Message 1978639.  

Hi David,
thank for this post,
Best Wishes,
Byron.
ID: 1978730 · Report as offensive

Message boards : Nebula : Machine learning in Nebula


 
©2024 University of California
 
SETI@home and Astropulse are funded by grants from the National Science Foundation, NASA, and donations from SETI@home volunteers. AstroPulse is funded in part by the NSF through grant AST-0307956.