[space.com] Mysterious Light Flashes Are Coming from Deep Space, and AI Just Found More of Them

ao_ika_red

Golden Member
Aug 11, 2016
1,679
715
136
https://www.space.com/41775-breakthrough-listen-fast-radio-bursts.html

Let me quote part of the news.
In the new study, Breakthrough Listen team members based at the University of California, Berkeley SETI (Search for Extraterrestrial Intelligence) Research Center applied machine-learning techniques to the August 2017 data set, which was acquired by the Green Bank Telescope in West Virginia and was originally analyzed using traditional methods.

The researchers, led by UC Berkeley doctoral student Gerry Zhang, trained an algorithm called a "convolutional neural network" to spot FRBs among the 400 terabytes of data. The strategy is similar to that employed by IT companies to optimize internet search results, Breakthrough Listen representatives said in a statement.

Zhang and his colleagues dug up an additional 72 light flashes, bringing the total number of FRBs detected on that day, from that single source (whatever it may be), to 93.
Please explain to me, if AI is already did a good job crunching enormous amount of data, why do they still need BOINC platform to do, what I think, the same thing AI already did?
 

lane42

Diamond Member
Sep 3, 2000
5,721
624
126
Well, imo its just another tool to help find E.T. Boinc isn't the only project trying to find E.T. iam sure.
The Boinc platform maybe a little dated, but still serves a purpose.
 

Tejas_2

Member
Jun 8, 2018
31
43
61
My understanding from the article is that they re-ran old data with different parameters and found the new signals.
 

StefanR5R

Elite Member
Dec 10, 2016
5,895
8,778
136
Buzzword BINGO! :p
Breakthrough Listen representatives said in a statement
This statement refers to "AI" only in its headline. I suppose PR edited it in, after they weren't entirely satisfied with the material which they got from the scientists.
Combing through 400 TB of data
Whoops, PR missed to add "Big Data" somewhere. Or is Big Data out of fashion already?
 

TennesseeTony

Elite Member
Aug 2, 2003
4,234
3,666
136
www.google.com
A thoroughly fascinating thread! It reminds me of the female scientist who was routinely combing through Hubble's secondary data, from the little onboard telescopes that lock onto stars to keep Hubble pointed in the right direction and found.....uhm....was it the ice geysers on the distant moons? or some more of the distant objects in the K-something belt (Kuiper belt/Ort cloud), way out past Pluto where the comets come from?

Anyway, pretty neat that just because 'old' data has already been examined using one method, doesn't mean that all the pertinent info was gleamed from that dataset. :) Personally, I think the FRB's were from an inter-planetary war, with nukes going off! ;)
 

ao_ika_red

Golden Member
Aug 11, 2016
1,679
715
136
Buzzword BINGO! :p

This statement refers to "AI" only in its headline. I suppose PR edited it in, after they weren't entirely satisfied with the material which they got from the scientists.

Whoops, PR missed to add "Big Data" somewhere. Or is Big Data out of fashion already?
Ah, got it. So it's more PR stunt than anything else. But it makes me wonder, if the researchers mumbled about "convolutional neural network", did that mean we would see the move from FP32 to FP16 workload? Because I think FP16 is pretty popular in machine learning space.

A thoroughly fascinating thread! It reminds me of the female scientist who was routinely combing through Hubble's secondary data, from the little onboard telescopes that lock onto stars to keep Hubble pointed in the right direction and found.....uhm....was it the ice geysers on the distant moons? or some more of the distant objects in the K-something belt (Kuiper belt/Ort cloud), way out past Pluto where the comets come from?
About time SETI add new sub-project. isn't it?
Anyway, pretty neat that just because 'old' data has already been examined using one method, doesn't mean that all the pertinent info was gleamed from that dataset. :) Personally, I think the FRB's were from an inter-planetary war, with nukes going off! ;)
Well, if that's the case, I don't really want to have contact with them. Let's imagine, if they already had interplanetary nuke war 3 billion years ago, their tech would be more advance than us and if they found out our existence, it would be "Battleship (2012)" or "Battle: Los Angeles (2011)" all over again. :eek::eek::eek:
 

StefanR5R

Elite Member
Dec 10, 2016
5,895
8,778
136
So it's more PR stunt than anything else.
Well, the discovery as such, as well as the apparent improvement of the toolset which enabled it, stand on their own, I think. I just found labeling a neural network an artificial intelligence a bit overblown.

Here is a nice site http://seti.berkeley.edu/frb-machine/ (some of you may have seen it already) which expands on the details of the discovery and tools, but also has nice background art which takes the AI theme a step further. :-)

From the paper:
Our technique combines neural network detection with dedispersion verification. For the current application we demonstrate its advantage over a traditional brute-force dedispersion algorithm in terms of higher sensitivity, lower false positive rates, and faster computational speed.
[...]
Our neural network is capable of processing Breakthrough Listen spectral-temporal data 70 times faster than real time on a single GPU, though processing speed in other contexts depends on the frequency and time resolution. We do not claim our technique is ready to replace current state of the art dedispersion pipelines, but our method shows advantage in some scenarios and encourages further exploration.
The training of our TensorFlow model takes roughly 20 hours on a Nvidia Titan Xp GPU.
[...]
Inference speed is crucial in real time applications such as autonomous driving, where large number of images must be processed per second. For radio astronomy, inference speed is less of an issue, [...] Our current model, without any inference acceleration techniques is capable of processing around 800 images per second on a Nvidia gtx1080 GPU, equivalent of around 70 seconds of observation.
In comparison, computations of HEIMDALL scales with number of DM trials. A search with 1200 DM trials process 3 seconds of observation per second. Thus in this setup our network inference appears 20 times faster than a brute force dedispersion. However, the different nature of the two algorithms makes it difficult to compare without ambiguity. [...] the main takeaway, that is neural network inference is likely more than fast enough for real time applications.
 
Last edited: