We don't have such a document unfortunately. Scott gave a good presentation on this at the workshop, including the details. That video should be up some time in the next few weeks.
Your second question is a good one. It is something we've struggled with as well. Some companies have internally developed algorithms, but they don't release them. You could look at Skyline, which is a pretty decent codebase for anomaly detection. We have also started an effort to create a good benchmark for streaming anomaly detection. The current repo is here: https://github.com/numenta/NAB We discussed this briefly at the workshop. We are actively looking for people to help us with NAB. In particular we want to collect a lot more data and finish implementing the scoring mechanisms. If you want to participate that would be great. --Subutai On Tue, Oct 21, 2014 at 8:44 AM, Nicholas Mitri <[email protected]> wrote: > Thanks for the tip Subutai! > The wiki page I’m reading doesn’t go into anomaly likelihood. Would you > happen to have a document similar to the one you posted about the CLA > classifier that I can dig more into for the mathematical formulation? > Something that reflects the work done in : > > https://github.com/numenta/nupic/blob/b6e5cf3c566e2d6ec60aeae24c4da4db27744138/nupic/algorithms/anomaly.py > > I’d also be interested in what algorithms you think would be suitable to > test HTM’s temporal anomaly detection against. I’m unfamiliar in my > research with any alg that’s comparable. > > thanks, > Nick > > > On Oct 21, 2014, at 6:15 PM, Subutai Ahmad <[email protected]> wrote: > > Hi Nick, > > At Numenta we use the difference between predicted and active columns, > plus the anomaly likelihood calculations. We've had very good results with > that combination. As mentioned on the wiki page when we tried out > confidences (over a year ago) we didn't get good results. > > --Subutai > > On Tue, Oct 21, 2014 at 1:59 AM, Nicholas Mitri <[email protected]> > wrote: > >> Thanks Scott. >> For the temporal anomaly detector, the wiki mentions using confidence >> parameters to calculate the anomaly score but the actual code uses >> predictive states instead. Is the latter the final approach Nupic is going >> with? Or should I be looking into reintroducing confidence based anomaly >> scores? >> >> thanks, >> Nick >> >> On Oct 21, 2014, at 2:21 AM, Scott Purdy <[email protected]> wrote: >> >> The algorithms are pretty geared around temporal data. If you have purely >> spatial data like your chart then I wouldn't recommend using NuPIC. You >> could use the spatial pooler and use the average overlap of active columns >> with the input bits to approximate it if you really wanted to use NuPIC. >> >> On Sun, Oct 19, 2014 at 12:28 PM, Nicholas Mitri <[email protected]> >> wrote: >> >>> Hey all, >>> >>> I was just reading the anomaly page on the wiki and was curious if >>> there’s an implementation of the non-temporal anomaly detection. >>> I’m running an older build of nupic and I can’t seem to find an >>> anomaly.py file like the one available in the current codebase. >>> >>> I’d like to try it out against other spatial anomaly detectors (euc, >>> manhattan, mahalanobis, etc) and see what kind of boundary it creates in a >>> 2D feature space. >>> The image below is the result of using 1-class SVM as a novelty detector >>> (from scikit-learn tutorials). I’d like to investigate what kind of >>> visualization the spatial pooler and the non-temporal detector would >>> produce. >>> >>> <figure_1.png> >>> >>> best, >>> Nick >>> >> >> >> > >
