Numenta NuPIC Tutorials, 7 Oct 2015

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Do you need to swarm for anomaly models?

Description

This discussion is a response to a mailing list conversation: http://lists.numenta.org/pipermail/nupic_lists.numenta.org/2015-September/011830.html

A

Just let it go so. Here's here's the question that Jonathan Mackenzie has on the mailing list he's trying to set up his model params. It wants to do not only detection. He wants to do it on multiple inputs coming in at once and use the swarm, but the swarm only creates it only works against a pretty good feel like it tries to get the best model or predicted to do leave. It's.

B

Not really optimizing.

A

For anomaly detection at all: what's the best avenue of attack there, you could just do a swarm pick. One predicted field go back and encode all of.

B

A

Then, and then turn it into a temporal anomaly right is that.

B

What you could do, yeah.

A

That's a good question: I.

B

Think this comes up a lot with anomaly. Detection problem is: where do you have some set of fields, and you want to know what something usual is happening and I.

A

B

Really have a clear prediction manager, so I.

A

Think the simplest thing to do in this particular case.

B

Is probably just sort of run this warm for a specific field, as it predict the field that should give you a sense of what are.

A

Some other physical parameters.

B

Are and then you.

A

B

Just do so I mean on the one end. You can try and you'll experimenting. So you can add all your fields back in right. Anybody like.

A

The encoder configurations for every field, yeah.

B

Model parameters- yeah one thing: that's a long hard now is the swarm- is my unit drop, almost all the fields, yeah.

A

When you're swimming on one of them.

B

A

Because most of the information.

B

That it needs for predicting that field comes from that field right and if there are other fields that are relevant, is probably not like small relative to end about you, physical.

A

Field, that's just for prediction. The anomaly detection function does look at all of the encoded.

B

Data not just a pretty real, so.

A

It said it is something to try which is art. If this go in there and add.

B

Iterations for all the.

A

Coders long as the other min and Max correct for the input data it might help yeah.

B

And a lot of times when you have a bunch of feels like this.

A

B

Like dr feels, like they're different thickness events for some field show and.

A

You had an encoder.

B

For that type of data, you might just have a single field. That's the combination of all the same things together. Could.

A

You still the problem of.

B

How to swarm it's.

A

B

Easy answer that well.

A

For anomaly detection, you know we use a certain set of model parameters where all of our are scalar anomaly: detection models. You could just use those parameters and have just slamming the min and Max reach encoder and just go with that and not strong at all. But maybe we.

B

Yeah, the only thing should you say and put all this yields it yeah.

A

Well, as many as you have to manly look at them and decide this even been a contributor in his case.

B

He's looking at traffic sensor so as long as they're.

A

Just data for a complete, potentially contribute.

B

To a comment, I think that would be totally reasonable just to take some parameters that work well with some no data analyst on add in all your fields and see what you get yeah in his particular example. I noticed that the results he was, they didn't tag bad pretty bad parameters. For instance, if you look at the the scalar field for field 152 mm.

A

B

N is 22 and the W is 21, which means they'll only be two buckets. The only.

A

Two possible.

B

Representations but.

A

B

A smart came up with.

A

That right, so that's something that so.

B

A

B

Going to realize.

A

That that's a problem, this.

B

One thinking, maybe it's better.

A

To just use our canned anomaly detection model parameters that we used for a proc for ITV yeah,.

B

I, don't think there's going to be an easy way to swim, though, on this particular dataset right there in trying to see so he's saying that that is the predicted guilt-filled.

A

122, it's a pretty good field.

B

Now, I, don't think he shared the data with us but I'm, guessing that this is either a field. That's not predictable at all, so just having two different IP stations, it's just as good as trying to represent it. Yeah.

A

B

Accurately, the.

A

Swarm results will print the nerve and their sports.

B

It's also possible that this field actually only had two values, and so this is totally fine. Yeah.

A

I, don't have the raw data, you gave me a Sam of it, you could look at, but it's it's.

B

Not a public data set but trying to swarm over all these parameters for all these fields. It's just not gonna work.

A

Yeah, so I think it might be best to try and start from our arc and anomaly detection parameters, but then make sure that each encoder has the right minute max and look at the data streams with us and make sure that they're valid data streams right and just.

B

And run it and see what the.

A

Anomalies, look like I mean you should have identifying your data set, a point where you know something anomalous happened and see if a.

B

New pic is showing the same I think that's probably a good next time to.

A

Just take the data include all the fields with parameters from a gem, yeah and.

B

See what you get all right, thanks, Scott thanks man.