Hi there!
I have seen that there are some examples of the application of NuPIC with EEG data. I wonder if there is any example with htm.core. It would be even better if it was on seizure prediction.
If there was not, how reproducible are the examples on NuPIC with htm.core?
Thank you in advance
Hi @hrodrigon, welcome!
I’m not sure about the EEG applications specifically, but I have been able to reproduce NuPIC’s NAB scores with htm.core.
Do you have access to any EEG data with known seizures? If so I could run it thru htm.core and see if the anomalies serve to warn about imminent seizures.
Hi!
Thanks for the response.
That’s nice, could you please describe a bit detailed the procedure you followed to reproduce the NuPIC’s NAB scores with HTM, I would like to try to do something similar.
As far the EEG data, I am still seeking for a good database for seizure prediction, whenever I find it I can make you know.
Thanks again.
Yea that’d be great!
So I created a wrapper for htm.core, which executes the core HTM functions in order (encoding → SP → TM → anomalyscore → anomalylikelihood). A main function takes in a data set (like those in NAB), initializes a model using given hyperparams, trains the model on given data and returns the HTM anomaly scores in the requisite format for NAB’s scoring functions.
An outer loops calls this HTM function on all NAB datasets to yield the anomaly scores, then I call the NAB scoring function (NAB/run.py --score --optimize --normalize) and see the standard score around 65, as with other HTM implementation including NumentaTM (NuPIC) & htmjava.
This is the high level description. If you’d like to delve in more detail it’d be easier on a video call.
Hope this helps!