Numenta 2015 Spring NuPIC Hackathon, 8 Jun 2015

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Blink Detection [DEMO #3] (2015 Spring NuPIC Hackathon)

Description

Using NuPIC to detect blinks in Muse data.

http://nupic2015spring.challengepost.com/submissions/37834-blink-detection

A

There we go hi, I'm Paul, that's George, we're both from this company called interacts on. We make these brain sensing EEG devices called muse, and we so we had the goal of being able to identify blanks with an h GM as good as our own internal algorithm that we use to identify blinks.

A

So the way in EG device works is that it detects your brain signal as long as you're not moving, and if you move like blink your eyes or something that muscle movement overwhelms the brain signal, and so you can easily pick it out with your eyes. So we thought this would be a pretty easy test for an HTM we're just trying to get up to speed on HTM and how it works and how to program with it.

A

So Muse data is an EEG device is also basically a voltmeter. So it's like just a wave of data. There's four channels in this we decided to try to use a single channel, which was the left ear, to try to be able to check blinks. So, first of all give you idea of what the typical data looks like. The blue line here would be your EG and then the red line is our own algorithm that we use to detect blinks.

A

So you can see a blink here goes down, it goes up and it goes back to normals and that's how you can visually detective Lynk and we're hoping to use use our existing algorithm to train the new HTML girdim.

A

So you can see all these all these red spikes here are our existing algorithm, detecting a blink.

A

Ok and then so we don't really have much to show except. Oh it's done. Ok, so we were try. The biggest problem we had was to try to get the data small enough so that we could actually process the data in time to do something useful. So we kept shrinking down the input data because miu samples at 220 hertz per second, and so we we downsampled to average over 10 samples and our this is on the on the left side. Is the HTM predicting whether you blinked and on the right side, is our existing algorithm?

A

Oh, it's vice versa. Okay, reverse what I just said. So this is HTM. This is our existing algorithm, so it seems to be doing pretty. Well, it's only miss predicting occasionally this is so. This is so here we have some mispredictions, but most the time it seems to be working pretty well yeah. So that's what we got do you want to have anything yeah. Any questions, yeah well, good,.

B

Predicted stuff, oh yeah,.

C

B

C

D

Okay, well, you know when you have three predictions of a blink in a row, how far part of those separated in time.

A

Yeah, that's that's a single blink right so.

D

That you, you were saying in this case you say: oh, we missed a couple, hm it's a couple, but a single.

A

D

It didn't really miss it.

A

Exactly right, the only time there's only one instance where it completely went wrong. Where is it.

C

A

There's one instance where it mispredicted.

D

And how are you making the prediction with the HTM imeem you? It builds a temporal model, the data, but what are you doing to? What exactly are you doing to make the prediction.

D

Using our prediction: what.

E

Do we have is this: is a classification task son? It's not really a prediction: I'm trying to figure out how you said.

A

Okay, so yeah we talked about yeah internally, we talked about whether to use prediction or anomaly detection or what, since it's a binary value like you, blanked or didn't blink, it kind of can fit into prediction because you base it on the EEG data and you predict basically because we had a way to classify the links already it didn't have to. We.

E

Have to figure that out, like I guess.

A

We had labeled data, alright.

E

What exactly? What is the file? Look like that you're feeding into the HTM you.

C

Don't prediction.

A

Looks like this, so we have voltages and then whether a blink happened and.

E

And both are being so, but both of those are being fed into the HTM and it's predicting the next time step exactly.

A

E

Ok, so the right column is also being fed into the HTM.

E

A

Right column is what it's supposed to be using to predict the left. Sorry, the left column was supposed to be using predictor right, column, yeah.

E

But it doesn't sound like it's set up that way.

A

It is that the maybe are just our output doesn't.

C

A

It, but it is set up that way. We think.

E

A

E

Have an existing code sample that does that. So that's why I'm curious how you set it up that way. Well,.

A

We talked to a lot.

E

Of people can talk, we can talk offline.

E

Then I understand the task. Ok,.

E

Not to predict forward in time.

E

The identity function are something that yeah.

B

This so that different interpretation of what this is showing from my understanding. So here we have and.

E

Can you speak into the mic a little here.

B

We have a current value that the system is seeing, and here we have a potential predicted value. So this is the blink information that it's seeing and the right column it's what it is predicting to see so right now it sees one next one. It predicts to see one as well.

B

Here it is one predicts that there will be no bling further down. Yes, it is.

A

D

D