Andrew, it looks like Mario's data has a timestamp in the first column, although I don't think he needs one. This data could probably just be input without encoding a datetime, because there are likely no hourly or daily patterns in this subset of data.
Mario, any chance we can see your code? You haven't called "disableLearning()" on the model instance, have you? If you can push your code to an online repository, I'll try to run it. There is definitely something wrong here. NuPIC should be learning those patterns and not returning high anomalies. It's probably something silly, but hard to know without seeing the whole code with the data and params. Have you tried using the AnomalyLiklihood instead of the anomaly score? See https://github.com/numenta/nupic/tree/master/examples/opf/clients/hotgym/anomaly/one_gym#anomaly-likelihood for an example. --------- Matt Taylor OS Community Flag-Bearer Numenta On Sat, Nov 1, 2014 at 3:21 PM, Andrew Currie <[email protected]> wrote: > Hi Mario, > > I am working on EEG data and anomaly likelihood and had a look at your > info. How do you encode the time stamp and feed it into learning anomaly > calcs? I could not see any time stamp encoding. > > Andrew > On 2 Nov 2014 03:48, "Mario Tambos" <[email protected]> wrote: > >> Hi Matt, >> >> thanks for answering. >> >> >>> From the chart you showed, it doesn't look like it gets >>> >> up to 1.0 very often, especially after it sees more data. In that chart, >>> how many rows of data has it seen? >> >> >> At the start of the plot, the model has seen 150000 rows. >> >> Do the scores get lower if you run it longer? >>> >> >> Nope, it just keps jumping every time it sees a QRS complex. >> The problem is actually stability. >> I thout that, after a time, the CLA would start returning generally low >> anomaly_likelihoods, but this doesn't happen. >> That "waving" motion (meaning the QRS peaks go up and down) you can see >> in the plot is constant though, could it be that that's the cause why the >> CLA cannot build a more stable model? >> >> Thanks, >> >> Mario >> >
