4 ms·

Machine Learning Predicts Laboratory Earthquakes

22 points by sizzle 9y ago

tw1010 9y ago
Let's just fill HN with every article on statistics there is, why don't we? I'll start with the stuff from 100 years ago.
- jacquesm 9y ago
  First published: 22 September 2017
- castle-bravo 9y ago
  Machine learning, bro. The models have way more parameters than we can possibly justify. Our predictions are meaningless outsisde the extrema of our training set. It's the future. It's a /revolution/ in AI. It's so far beyond statistics that it's not even the science anymore.
dane-pgp 9y ago
Why look at "laboratory earthquakes" when we have plenty of data about actual earthquakes? We sort of accept that after a big earthquake, there is likely to be smaller earthquakes nearby and a short time later. By looking at all the available data on all the major and minor earthquakes in a region over time, I'm sure there would be other patterns found, even if the predictive power is not actionable.
- candiodari 9y ago
  Really ? Where can I download long term acoustic data from earthquake fault lines ? Needs to be years, and ideally including more than a few real earthquakes (let's say 50 to 100 of them would be ideal).
  - dane-pgp 9y ago
    I didn't mean acoustic data, but I was thinking of long term global datasets like this: http://www.isc.ac.uk/iscgem/overview.php http://www.isc.ac.uk/iscgem/overview.php It covers over 26,000 earthquakes going back roughly a century, but I admit I'm somewhat disappointed that it only seems to cover Magnitude Mw ≥ 5.5 quakes. A machine learning algorithm may need even more data points than that to spot reasonable patterns, even if it was given the location of landmasses and fault lines to help it, but I wouldn't be surprised if someone could generate a model which can predict future quakes at least somewhat better than random chance (factoring in averages for a given location).
    - candiodari 9y ago
      That's just locations and quakes. And even there, only really big ones. That's only half of what you need to predict quakes. There's not much in that dataset to predict with.
thisisit 9y ago
So predicting earth in a controlled environment, something which might have been generated using a pseudo random generator code? How will this hold up in nature?

22 points by sizzle 9y ago

Let's just fill HN with every article on statistics there is, why don't we? I'll start with the stuff from 100 years ago.

jacquesm 9y ago
First published: 22 September 2017
castle-bravo 9y ago
Machine learning, bro. The models have way more parameters than we can possibly justify. Our predictions are meaningless outsisde the extrema of our training set. It's the future. It's a /revolution/ in AI. It's so far beyond statistics that it's not even the science anymore.

dane-pgp 9y ago

Why look at "laboratory earthquakes" when we have plenty of data about actual earthquakes? We sort of accept that after a big earthquake, there is likely to be smaller earthquakes nearby and a short time later. By looking at all the available data on all the major and minor earthquakes in a region over time, I'm sure there would be other patterns found, even if the predictive power is not actionable.

candiodari 9y ago
Really ? Where can I download long term acoustic data from earthquake fault lines ? Needs to be years, and ideally including more than a few real earthquakes (let's say 50 to 100 of them would be ideal).
- dane-pgp 9y ago
  I didn't mean acoustic data, but I was thinking of long term global datasets like this: http://www.isc.ac.uk/iscgem/overview.php http://www.isc.ac.uk/iscgem/overview.php It covers over 26,000 earthquakes going back roughly a century, but I admit I'm somewhat disappointed that it only seems to cover Magnitude Mw ≥ 5.5 quakes. A machine learning algorithm may need even more data points than that to spot reasonable patterns, even if it was given the location of landmasses and fault lines to help it, but I wouldn't be surprised if someone could generate a model which can predict future quakes at least somewhat better than random chance (factoring in averages for a given location).
  - candiodari 9y ago
    That's just locations and quakes. And even there, only really big ones. That's only half of what you need to predict quakes. There's not much in that dataset to predict with.

thisisit 9y ago

So predicting earth in a controlled environment, something which might have been generated using a pseudo random generator code? How will this hold up in nature?