Thanks Matt :).

The fact that it predicts the last value it saw when confused explains a lot! 
I’ll definitely look out for Subutai’s (timely!) presentation. 

best,
Nick

On Apr 26, 2014, at 6:48 PM, Matthew Taylor <[email protected]> wrote:

> Comments inline...
> 
> On Sat, Apr 26, 2014 at 6:30 AM, Nicholas Mitri <[email protected]> wrote:
> Hello all,
> 
> I’m a little confused by how Input/prediction pairs are formed and plotted in 
> some of the tutorials. I’ve attached a 1 step prediction plot for the sine 
> wave tutorial as well as 1 step and 5 step prediction plots for an apnea 
> predictor I’m working on for reference. 
> My understanding is that the number of steps specified for multi step 
> prediction dictates the delay used to associate an input with the proper 
> prediction. So, following the sine wave tutorial, part of the code includes a 
> shifter class that (I assume) queues the predictions until they’re needed. 
> 
> I'm not exactly sure how the shifter works, but you're probably right. 
>  
> Looking at the sine plot attached, I don’t understand the shifted plots i.e. 
> the lag in the predicted plot.
> 
> While the prediction is usually lagging, there are times when it is not, 
> which means the shifting is occurring properly. The results don't look that 
> usual to me, because you're only a few hundred rows into the data. It takes 
> quite awhile for NuPIC to get better at predicting sines (it's not very good 
> at it anyway). 
>  
> Also, the anomaly score seems erratic. It’s 1 between samples 60 and 100 when 
> the 2 plots are identical, which is weird in itself since the model hasn’t 
> had enough time to learn the bottom portion of the sine wave. 
> 
> This doesn't seem unusual to me either. During the time when the score is a 
> flat 1.0, it is continuously wrong with its prediction, simply predicting the 
> last value it saw (that's what is usually does when it doesn't understand any 
> patterns). So an anomaly score of 1.0 during this period makes sense because 
> every value it sees next is misunderstood. 
>  
> 
> The same applies to the other attached plots where anomaly is max in segments 
> where prediction and input are in line, low in segments where they are off, 
> and 0 for very low input values which is misleading. 
> 
> I’d appreciate some guidance in analyzing these plots :).
> 
> You know who knows a lot about anomaly scores? ;) Subutai Ahmad. Luckily, 
> he'll be giving a presentation about them [1] at the hackathon next weekend, 
> and I'll be streaming it live on UStream. 
> 
> [1] 
> http://numenta.org/hack/schedule/#anomaly_detection_using_the_cortical_learning_algorithm_subutai_ahmad
> ---------
> Matt Taylor
> OS Community Flag-Bearer
> Numenta
> 
> _______________________________________________
> nupic mailing list
> [email protected]
> http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org

_______________________________________________
nupic mailing list
[email protected]
http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org

Reply via email to