Hi Nick, Note that NAB does not do classification - it just does anomaly detection in real time scenarios. I think temporal classification with HTMs is an area of research (we haven't done much with classification at Numenta and unfortunately we don't have clear worked out examples).
--Subutai On Sun, Mar 1, 2015 at 9:26 AM, Nicholas Mitri <[email protected]> wrote: > Thanks Subutai! You mentioned the project to me in a pervious exchange but > I couldn’t find it. > I’ll have a look at the code. > HTM didn’t perform too well in a set of classification experiments I > prepared for my thesis so I really wanted to showcase it doing what it does > best. Unfortunately, I only had the hot gym example which I don’t think > offers enough insight into how HTM can be put to the best use. Hopefully, > the NAB project can offer a better alternative. > > Thanks again, > Nick > > > On Mar 1, 2015, at 7:18 PM, Subutai Ahmad <[email protected]> wrote: > > Hi Nick, > > Are you referring to NAB (Numenta Anomaly Benchmark)? The code for it is > here: > > https://github.com/numenta/NAB > > The purpose of NAB is to establish a benchmark for real time anomaly > detection. One of the goals is to include actual real-world sensor data > with labeled anomalies. We’re at an “alpha” stage right now so it is not > fully complete but you can look through it (there’s a doc on the wiki). > > NAB is not specific to Numenta. We’ve included Skyline (a popular open > source anomaly detection algorithm) but we hope over time people will add > other algorithms. > > It’s not fully ready yet but I’m happy to go over details in the next > office hour if there is interest. We could really use help from anyone who > can provide real sensor/machine data with anomalies. > > —Subutai > > > > On Fri, Feb 27, 2015 at 3:31 PM, Nicholas Mitri < > [email protected]> wrote: > >> Hey all, >> >> There was talk of an ongoing project to benchmark HTM against a set of >> algorithms. >> Any updates on that? >> >> I’d be interested to see what algorithms the Numenta team finds >> comparable to HTM. >> >> best, >> Nick > > > >
