Numenta Featured Applications, 22 Oct 2014

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: A Whole New World [DEMO #4] (2014 Fall NuPIC Hackathon)

Description

By Chetan Surpur & Yuiwei Cui.

A demonstration of sensorimotor inference in simple robotics.

A

Hello, uh I'm trading, this is ua and our hack is a an application of the recent research we've been doing at dementa regarding sensory motor inference and temporal pooling.

A

I will show you a little robotics based hack uh on that.

B

Yeah, so this is actually the first thing: uh the first experiment. We use real world data and tested algorithms, so the setup is very simple: we have a smart driver. There has ir sensor, that's taking uh the distance between some objects.

B

It has a range between about 5 centimeters to up to 30 40 centimeters, so it also has a motor on the back. If you only look at the sensory part, it looks random because I reprogrammed it to move randomly to sample a fixed motherboard, but if you consider both the sensory input and the motor command, it contains some information about the layout of the world.

B

And, if you also, if you change the spatial configuration of the work you might get, a sense of this is a new environment versus the old one, um and generally this kind of algorithm, yeah.

A

So, a little bit of the theory um before we show you the demo, so you kind of understand, what's going on uh what we're doing is taking and taking this data from the sensor and the motor and feeding it in using a scale encoder to get sdr's and feeding it in both of them concatenated to layer four, and so, if you think about it from the perspective of this robot, what it's seeing is it thinks okay, I'm gonna move left now and then it moves left and it senses how far away the object is using its ir sensor and then it can turn right.

A

I say I'm going to turn right now. What do I expect to see how far away the object is it's supposed to be and it can make a prediction. So in this case, layer 4 gets the information about the current sensor. Reading how far the object is looking at is and the motor commands about to execute and layer. Four basically learns those sensory motor transitions and learns to predict what it's going to see. Next, what the sensor value is going to read next is going to look like.

A

So if this was predicted by layer four, if a transition was predicted by layer four successfully to learn that transition, then uh now there two three can pull over those predictive transitions. Because now the world is more predictable, it can build a stable representation for that world. So layer, two three pulls over. It does temporal pooling and if it was unpredicted, then this uh then you'll see bursting in layer four and uh it'll pass through those changes.

A

So what we hope to see is some uh stable representation once the world becomes predictable, so and layered, two threes is supposed to learn: high order transitions, but we didn't test that part okay. So, let's take a look at the demo uh disclaimer. This is a live robotics demo, so very likely it won't work.

A

It did work yesterday uh in the room and we took videos of it but uh go ahead and stand up. So you can see all.

B

A

So there are three objects.

B

This is a robot, the sensor is in the front, so you can sense the distance to the object and it will move around to sample the three objects.

A

Movements are random, it's.

B

Random, it's movements are random, except.

A

uh It's with that.

B

It tries to explore.

A

Transitions, it hasn't seen before so.

B

It's biased, exploring new things.

A

Okay, uh it's initializing right now. So what I'll do is.

B

So we want to look up there or we want to look over.

B

Here, okay, so we just tell you to do 30 random moments.

A

30 random movements- this is by the way the nothing has been trained. Yet the model is empty. Hasn't learned anything at this point, so here what you're seeing is uh the representation in layer 4 in the middle representation in layer 3 at the top and the number of unpredicted cells in layer 4 at the bottom. In that graph, there.

B

Everything's unpredicted.

A

It starts to make some predictions as you can see in the bottom graph, there you'll see in the top layer. It made a sound and that's it classified successfully this world, you have to see a stable representation in layer.

B

3., this is only showing a stop sample of the number of columns, so.

A

Yeah, this is just a subset of the number of columns, so you see some stability there so now in there and for this world, if there's a stable representation, it's built because therefore was able to predict to a different.

A

B

Yeah, I just changed the spatial configuration so.

A

It looks it should look yeah yeah. We mark.

B

The lines so that we can go back to the to the first word.

A

B

We'll do in a second.

A

But it looks, it looks different, we'll see what it um what it sees and what it represents.

A

B

A

See that it's more everything is unpredicted again, because it hasn't seen this before layer. Three there's there's no stability. It keeps changing between the representations um and soon you see that it starts making predictions in layer four and.

B

It recognizes this as a new world, and you see this representation.

A

Here is stable right, but it's different from this representation for the previous world, and you see this classification here. It used to say zero for the previous world. Now it says one and it also plays a different, sound.

B

Now maybe I'll go back to the first word to see whether it can recall all right. Are you sure that that's exactly the first word? We try to mark the the positions yeah.

A

There will be some small variations of noise in the sensors, the motors and in our movements here, but hopefully it can lead past that it was this.

B

A

Okay, so let's see what it sees.

A

It's not perfect, so yeah, so the prediction says: there's noise right! It's the predictions are some of the columns are still bursting yeah.

B

Is that? Because it's just slightly different, slightly they're finished after the noise in our setup.

A

Is a new system.

B

Now you'll see up there right there see this was the representation I had for world.

A

Zero, this is the representation it.

B

Now sees it's interesting, so it's.

A

Somehow generalized a little bit accepted this noise and so yeah. I couldn't consider it didn't think of it as a new world that this isn't.

B

B

What is this new reality.

A

So some some of the is saying, there's some predictions you can see, but also some unprediction, I'm pretty right.

B

And you'll see there's.

A

No stable representation in layer three, um but I should have maybe done more time steps. Do you think it would lauren has a new world.

B

Yeah, but I think I didn't.

A

Actually allow it give it enough time.

A

But yeah close zero just needed a couple, more steps, all right. Let's just.

A

A

But yeah, essentially you see that with the anomalous world uh it wasn't able to consistently predict so it didn't.

A

For some reason, so that's not as expected, but.

B

The review is the first.

B

A

And distinction in layer three and that's where we're going all right.

A

B

B