HTM as intelligent seismic detector


We are doing research on automatic earthquake detection. We tested HTM as a possible intelligent unsupervised approach and obtained interesting/intriguing results. Here’s a report of our tests:

arXiv .

If you have any suggestions, ideas or comments that you think it may be useful to us, you are welcome.
Thanks !

R. M.

Below a brief background of our research:

It is not so difficult to make a machine understand what is a big earthquake, but the task becomes very hard for smaller ones.

Small local jitters, slammed doors, a truck passing nearby etc etc are easily confused by most of the algorithms. We used in the past A.N.N., and trained them to distinguish instrumental noises from tiny seismic waves. This approach has its own advantages and disadvantages. We somehow succeeded to obtain decent results (and won two technology related prizes, if you read japanese Yahoo! Japan “Novelty Goods” prize and [Internet of Things prize].

In the arXiv report above, we changed approach and used for the first time NuPic HTM, with intriguing results (HTM adapts quickly to the instrumental background noise and recognizes very small deviations from it with high anomaly scores. Very encouraging as a first test). If you are interested and want to give us some comments/suggestions please read the arXiv report. THANKS !


Thanks for your work! I’m happy to see this:

After an initial transition time of about 200 thousands time
steps, the HTM cortical algorithm was able to consider the seismic sensor background
noise as normality, lowering its anomaly score to nearly zero. Waveforms instead were
recognized reliably with repeated spikes of high anomaly score.


(By the way, I moved this from #other-topics to #htm-theory:papers.


Thanks a lot to you ! We were following your videos and all the other stuff you made for the community. Thanks again, I know how hard it is to clarify complex things to other people… we do this for our students as well. But in your case everything is new and under development, more difficult !




Hi, we submitted a work about this to the next SAI conference in London and I just received notice that was accepted (^_^) /*

If some of you will be in London at the conference, we can have a chat there.

The title is “Using a hierarchical temporal memory cortical algorithm to detect seismic signals in noise", conference will be the 10-12 July 2018 in London, UK. This is the URL:

Cheers and maybe see some of you in person over there !

R. Micheletto