Numenta Live Streams, 10 Jul 2019

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Visiting Scholar Pranavan's unsupervised continuous learning project

Description

We have a visiting scholar Theivendiram Pranavan from National University of Singapore who'll be talking about his work on unsupervised continuous machine learning.

A

Good morning, everyone so today I'm going to present one piece of work that we are doing in anyways like parenting. So our ultimate aim is to like create a. We shall learning paradigm with brain learning principle, so we are like only focused on computer vision problems.

A

So, for the moment like, we have found like not found that we read through the literature and we found and finalized few principles. So the visual learning is continuous, so I feel there is no training or testing phase. So we have to learn continuously so so we have to deal with the labor one and the unlabeled one. So how we are going to deal with that and it favors the recent information. Suppose you read something yesterday suppose you read something like three years back so there's a difference between the recency like there's.

B

A temporal factor.

A

And it uses both top-down and bottom-up processing, so in the case of deep learning, they're just going from pictures to a decision, so it's technically bottom on, but actually there are like top-down guides as for Indian children, and it.

B

A

High level semantics so I can show you one.

A

Okay, so you can see these pictures. The first one is like an illusion. So there's an there's, a jungle, the ideas arise and all daily. So at one time you can see only once either you can see the old ad or the young lady. So if you take these two figures like we are giving equal semantics, like preference for these two images. But if you look at the second diagram, it's kind of some scratches, but if you look closely, you can see a talk in the middle. So when.

B

You see the dog.

A

Then it's difficult to unsee the dog so so.

C

In the first diagram we have to intuition about this is every time you show this tall picture before it's been flipped, so it's not the same.

C

See the dog iodine stain at this point understand is seeing certain there are generations.

A

We should learning distant recognition, that's five and we usually jump to conclusion. So we don't do the whole forward propagation thing in brain. So, for example, if I have a dog in my house, it has a build in its neck. So if I am hearing a sound in the corridor, I'll be expecting the tone. If it's not coming like it's sunny, if you think so, we.

C

A

To conclusions.

B

D

You're going to say, question against.

A

I'm showing you something you are recognizing, whether it's a car or something like that. This is that you're jumping here, like you.

C

Are not doing your bell.

C

A

It's it's not like. You have to take a clear decision right. Maybe you have to wait for some time to take the decision right, I mean in.

C

The case of a picture like you have China you're, making a false distinction about visual learning versus learning yeah. All these attributes apply to all kinds. All learning in the brain, yeah.

A

C

Jumps the conclusion on I'll- just maybe this but I- think that's an artificial distinction. I think it's if you think it falls under communion. It's this ultrasound and you recognize it have some ambiguity about it. Just like mine ambiguity, about the visual image and to.

D

Me they're both sort of dealing with ambiguity.

C

You might think it's a different sense and you're thinking about visions. Maybe that's why.

A

And these are the place of the principle like yourself publishing these actually spectrum, so I think there are overlaps, so maybe different wording. So the first one is learning by rewiring space representation than embodiment.

A

B

You take a look normal.

A

Machine learning problem, what we usually do is like we just put some data just divided into training and testing. In some cases we divide the training they tie into, like validation, the actual training data. So let me train a classifier and evaluate on the testing, so it's not going to work all the time. So we have to do like some learning with all these, like all the data that we have, so we are not going to get annotated data all the time, so there are lots of unlabeled data in there.

A

There are lots of unlabeled data, so we have to liberate the information from those data. So the question is how how we are going to handle the delayed feedback so I'll just.

C

A

To explain you how so the delay tree pick suppose we have a child, so I'm I'm, seeing the child do not like a couple of cars so different cause. The child doesn't know about these cars, but he may know about other vehicles. So.

B

He might think like this.

A

Some vehicle, but he doesn't know it's a car so down the timeline. Maybe he will keep that in memory down the timeline. If somebody gives some feedback with a new car, somebody is feeling something.

A

So somebody is telling this is a car so from this information that he is able to update he's learning by like labeling all these previous car. Maybe he will keep these cars in the memory for some time. If he doesn't get any feedback, you will just delete from the memory so how we are going to handle this delay feedback. So so, in a neural network setting like we tried one implementation, so I am going to explain.

C

Possible onion I mean so what I was talking about this of the nature of specializations? And what do you manage you to get specializations that your area or your space of objects is naturally separated by not overlapping, and so it seems to me the way we do this is that even if I know one gives me a label, anything I actually am able to do this self classification and very, very well. I. Don't need a label to do that in all the time, not all the time, but most the time someone will say.

C

That's a car I already know what cars I just don't have a label for it. So it's like it's not like I couldn't learn anything it's like the situation comes across. Sometimes we say oh no. This is the Toyota and here's how you can tell the other and I didn't know that before so actually,.

B

I just question.

C

About is it really the legacy property? It's not like, if you think about machine learning,.

C

To train the network, but with a embody, sparse, centric motor system, I, don't believe you Nina, you just you've learned. You should pacify the world already and you thought label on it.

A

C

A

Doesn't mean to be at lay buddies, it can be some auxiliary information, I.

C

Mean I guess the question: is it's not a time problem like I have to go back and redo something that's what the word delayed.

C

A

Maybe the wording might be wrong. The thing that I'm explaining is right. You.

B

Are learning something.

A

You don't know what actually they are. You are keeping those things in your memory, yeah down.

B

The timeline you.

A

Are getting some feedback about the similar Jake? How.

E

We are going to learn all the.

A

Things that you have in.

A

I'm going to explain the.

B

First part is just an image.

A

Classification program, so we have a noodle head, pursue any input, an image to get a vector of probabilities or just you know, if I X I, and so you, you train the baseline model using.

A

So for the evaluation data, what they are going to do is we are going to keep a bounded queue with the fixed size and we are going to store all these inputs and outputs as pairs in the bounded queue. So you put an image and you get the probability vector. We are going to store these X and X or Y I power in the queue so actually.

B

A

To keep like a bounded queue, so if the queue is full, we are going to do something yeah.

C

Using keeping all this.

A

B

Are not we don't know the name word for.

A

B

These are the results.

A

Produced produced by the network.

C

Probability of what buses bus.

E

Yeah, but did not be a label.

E

A

Not the actual, thank you. These are model, you just pass six and you get a probability vector. We are going to store in a bounded queue. If the queue is full, we are going to update the model even without the actual label. The queue is not full, but we are going to get the next X and we have adopted the monarchy. So if the queue is full how we are going to update the model?

A

B

A

You can remember the example I saw so we are going to give a feedback in the form of Z 0 and y 0 y.

B

A

The table, so we are going to say what is this object and we are going to pass some auxiliary information like you can pass whatever. In this case, we are going to pass an image so.

B

We are going to say this is a car.

A

Now it is show some car and you are going to check in the queue whether you have any cost.

A

B

Is the feedback that we're generating and.

A

Q so from this red 0, actually using some content-based image, retrieval like we can create some features so for sex 0 we can create 0 Prime and you have a queue of images. So for all these images, you can do the same thing. So now you have all these features, so you can find the similarity between these are 0 Prime and all the images in the queue.

C

Good question contest page may be achievable, sounds like what is it? How do you retrieve something content-wise, yeah.

A

So we actually use the features from there. So in this case we are using color histograms. You can use shape.

C

So something it could be more.

A

It could be wrong yeah, it could be wrong. You know.

F

B

Some training right yeah, you just doing.

F

Like some clustering based in some features, yes,.

A

F

Is not trained right, yeah.

A

We do know trade, so we have this.

B

Q of Avengers right so.

A

You can find the color histogram for these lineages cluster.

F

Based on those games so.

C

F

C

So the goal of the whole system is to be able to close with them properly. So isn't it I mean I didn't get.

C

Seems like the goal in the end, the goal the system is given image and you cluster it correctly. What does this belong to? But here you doing a very crude method of it, so it's likely to be wrong. There's no training involved. Yet right. Yeah I show you two pictures of cars unless they're really similar in some basic ways. It's not going to work right, yeah.

A

So, actually like we, we are going to use Z as well as the values products produced by the model as well. We.

B

Just not, we do not exactly depend.

A

On these colors Instagram or shape or textures, as we only use the values produced by the motor as well, so this something like that, so so.

B

We are going to check the distance between.

A

Z 0 Prime and chat, I prime, so these are like the input space, so we transform the input space into exam, 0, Prime and certified, and the other distance is the like the neighbors space. Actually, these are due probability distribution. So we can use some very clear.

B

A

That area to find the distance between the labor.

B

A

And for the distance between input space we can use, we are using color color histogram we can. Since we are using a smaller image, we are using color histogram. We can go for change picture as well, so what we are doing is if the distance between y 0 and y arranged close as well as the distance between Z 0 prime ends at I prime, is close. We are going to update by a very small learning rate, so it is an integer which is greater than 1.

A

So R is the original learning rate that we used to update the base model. So in the case like we think, like the model is learning correctly, so we are not analyzing the model with a higher learning rate and in the other case, the distance between the labor space is high. In that case, maybe they.

B

Are not learning.

A

It correctly, in that case, we are penalizing with the right higher learning case. So we are just multiply everything with the fact. Actually, these a and B can be learned. I should be used, some arbitrary values like 5 or 4, but actually a B can be learnt so in for all the other two cases we did not update.

A

That's what we have done. I can show some preliminary results as well. Any questions. What is the distance.

F

One can use so.

A

For the first one, the distance between labels, which we are using Bhattacharya Stan.

F

A

Distance, okay: you can use more distance, so the KL.

B

Divergence of any two.

A

You have to measure the distance between two probability distribution, so you can go by.

F

A

You have experiment.

F

E

A

You can use either chi-square of some correlation function, so there are like four or five different ways: I think in.

C

C

Train, oh yeah,.

A

I can design on that like to do something, but we did notice that, for you know where.

C

A

C

Okay again, that could be wrong right. That would be wrong right. No, no, it could be like hey. These are both cars, but they very very different. Yeah.

D

You could have a red car or.

C

D

I think in here there's no, nothing happens in that case. I think you're, probably relying on that there'll be some other training examples e0 some crumbs that that hopefully will match yeah.

A

Listen actually like even for the auxillary informations em, so you know like we don't have to put only for naming. Maybe we can put the batch of images so there's f0 optionally information that we are giving it as a feedback. So the problem will occur like if you are using a single image and if you can use a pasture of images like somehow we can.

C

What seems means that your user views in yourself untrained classification system here to do your training and, and so it's gonna, get it wrong, sometimes yeah. The question is: will it overtime fix those problems or not and how it's not, obviously how it would make it well? I I have.

F

A follow-up question: the.

B

Reason when we created.

F

Same so CNN's our automatic feature extractor, so you will have to use many features like we mostly do in computer vision. The.

B

Best, so why not use.

F

Like the latent representation, one of the first layers in the CNN.

A

Instead of using this, oh yeah, our features, okay, so actually, like ceilings, are learning or all these like shapes and colors, and everything in the form layers. Actually, they are learning. We told like this can give like some other aspect of the color. It's.

B

A

Learning everything actually, it's learning some um some aspect of color shape and the.

F

Main difference there is that it's learning right, I suppose to the other.

B

One, it's not learning it. Also, it's don't to fail at some point.

E

A

B

C

Don't actually.

A

Depend on color, you can use.

C

Whatever it is, untrained classifier, so it's going to get it wrong. I mean no.

A

Actually, the baseline model was trained on some label dataset, so it's not actually like you're.

C

Starting producing do you coming from to be you're, assuming that the new examples are gonna be somehow similar, these base levels, but.

C

B

The whole point.

F

Of continuous learning is that you have distributions which are constantly shifting.

A

Underneath, like maybe in the town, don't the time, then you'll get a correct label for wrong.

A

B

A

We did some testing on C for 10.

F

F

You expect the queue to be foo, and then you run one pickup date and then here so.

B

F

A

When the queue is full, we search for the best match. Okay, and we remove it. Okay, if we couldn't find a basement, we just remove the whole desponding the queue so.

F

It's not like frequent at every iteration. You.

A

Were doing like in a batch.

F

A

It will take some time to like fill up the guy.

F

A

The queue is full for every new sample. You have to update right, okay, so.

F

You use the same sample to update more than once. No only one is after.

B

F

Use it you just discard yeah.

A

F

Then your queue is empty again right, no.

A

F

Only want that on day, one.

E

A

B

Actually, like we have made it.

A

On site for ten, so actually let me use the like very small CNN with likes 60,000 parameter, so the baseline accuracy was like 65, so with various queue sizes we just evaluated the results actually like it's always possible, like this accuracy may go down in the wrong direction like, but in this case we tested their quality is increasing with the Q side. So if you are.

B

A

M's auxilary information, even with that it's increasing, but.

B

If you are using some.

A

Oxy dolly information. The accuracy is increasing to 76, so.

B

A

A possibility, like you, can this there's a possibility like this one can do mistakes, but for.

F

This tennis, it is in what is life without you. Topsy is just using the widest.

C

A

So, actually, using the test dataset to update the ways now.

D

By test accuracy, I think you mean the baseline model.

C

The original small.

D

Network that it was trained on before he does all this cues and.

C

Is this yours or someone else's now is.

A

There I want to go this small see it.

C

So that's what you got before you did any of those updating stuff.

C

And I think it's fair to say that today's standard on support end is much greater than this I mean yeah.

F

C

So we're really showing that the the effect of this train methodology, this particular model on.

D

A small model than all other state-of-the-art, normal.

A

State of the other side on like 96-97.

B

You know we want to see here like much.

A

B

A

It like small all the way and.

B

65 is the like highest accuracy.

A

Most achiever would be like the 50,000 be sparkles.

A

It's actually like we're running another set of experiments with amazement, so we higher number of sample, but.

B

We are not using the entire method.

A

We are using like one percentage of the day per samples.

A

And like we cannot compare with 60 5.82 because, like we.

B

Are updating for test dataset? So actually we are learning something.

A

From test test well, the baseline model is not learning anything from test data. So last time we pointed out a mistake, so so.

F

Continuous, like so, you have the full training set. You get one percent of the labels and then you are part about $2 to use all the remaining unlabeled image from the training set. Can't you do that yeah.

D

That was my suggested.

F

Here otherwise,.

B

I really do anything.

A

So our goal is like the current neural network models like if you take the memory size and number of parameter is not gross output. It's very tight and they're getting a high accuracy, so meet me down the timeline. We can create models like fewer parameters, and maybe even that close, it might be like slightly less I need to create.

C

So you're gonna to be high accuracy, but the relative additional parameters.

A

Should be closely as possible, but we may be like a 10-4.

A

Newman I am planning to do these things, so the first one is like they have published a paper in avi CMS of the public is sparse and burning.

A

Using CPC they are producing, these aren't super is burning so from using.

F

B

They are producing embeddings.

A

Which can be used to get like higher accuracy classifiers, so I am trying to use even for this CPC.

B

They are using neural.

A

Network with, like millions of parameters still there, they depend on like large.

B

Networks, I am I, am planning to use.

A

Fast networks to find these embeddings and compare the results, so we are.

D

A

I gave him the CPC.

D

Papers to look at he's got some code.

A

So even in CPC they they rely on, like very large networks like best net and obviously so.

B

Maybe like we can just replace.

F

You can try ami tiny'mon, I, think they're, 84 I think.

B

It's better than image net, like it.

F

Tonight and I have.

A

A GPU so I have all the image data, so I can just.

A

So maybe, during these three weeks, I should be able to produce some resource. I got the call right.

F

I still question on the queue: okay. So if I see like a blue card with that, okay and my first I say blue card with that, but I don't have to label and the fourth one I see is a blue card spot. So C is gonna match right. So I'm gonna use all those three samples on label to learn and then I'm gonna discard those three samples and Mike Weir is gonna, be able taking right. Yes, so.

A

It depends on your cue, like you like. That's why the accuracy is improving with the cue sighs, okay, but.

F

If I use all my samples, my career is gonna, be it's not always gonna before okay.

C

I'm trying to understand okay.

C

I'm trying to put that into a deeper understanding of what how that contrasts. Other approach. Isn't it's not obvious to me I'm, trying to sort through all these things to keep somehow organize my brain.

A

Actually, like most mutant it quite instantly, if you get the sample, you just update yeah.

C

A

This case we are keeping like cue samples actually did increases the context so, like not.

B

The single something.

A

C

Have why is that I mean if I'm training on every sample then I'm using all those samples? So what you saying, if I, have an excuse so now I'm able to get grouping training samples together that somehow that's better than training individually? Yes, so it's not obviously like you.

F

Can use a bad sample because he's scenario is that you don't have labels for everything so you're in a continues. Learning scenario.

B

You see a lot of things without and then sometimes.

F

Is it something new for labor? You know what it's saying that when you get the label, then you can search for all other examples are not seen about the label and then you, instead of learning just from this centrifuges and.

C

We're hoping that all those unlabeled samples that are similar, that they might be correctly yeah.

F

D

That's the key, there's no real unsupervised learning going on here. It's.

F

B

D

Finding a supervised label yeah.

E

D

I think but I think that's right, but and on top of that, you're also training the model on this new.

A

I think nobody else has done it for like label data, so I mean even for the normal training we can check like. We usually update the training learning rate by like over a box right. Maybe we start with the minus 4 there, like in minus five, something like that right. So actually.

B

If you are learning something correctly like we.

A

Don't have to update order with the higher learning rate track, so maybe.

B

A

First epoch, you can get the research right. You usually go like 20 30 bucks, a box right to get a country sauce.

E

A

Sorry speaks trial, even whatever they are. You have this holistic style. What it's saying is like you've got predicting. They will correctly.

F

Understand you saying saying the gradients are gonna be low. So, even though the learning rate is too high.

A

You won't get there.

A

F

Say there is still like small update and we can even I guess.

D

It a typical backdrop: you take the difference between the predicted and the actual label and it's some function of that. So the slope closer is just like you're saying, they're breaking and if this is like an additional factor on that, somehow.

A

F

Like it will be, like maybe cynical, adaptive learning rate models like Adam, so Aaron I also fixed that, because it's taking like the moving average of more green so and ends a way of fixing that it would be like an alternative right. That's what you're saying.

F

So you're gonna work on that.

B

D

Think right now, while he's here is gonna, focus on sparse networks and.

A

In your book, you always tell about life without a dealing with Tammy in all the computation machine problem. We are just just dealing with images right in your book. I read your book on internal. So, like you always tell like, without like dealing with temporal thing, we are not going to build intelligent machines right. Yes,.

B

Even I have the same.

A

Intuition so like do, you have any like workarounds did in that line.

C

D

We have a lot of papers.

C

C

Yeah that I work in sort of image classification is very, very recent. It's always been that this is a subset of inference. So my first comments here is: we've started here. I'm thinking about how the brain builds cell models is messed up. Mark is talking about redoing it yeah. That's what Marcus talked about where the the spatial inference is sort of the corner case of the general perception motor.

C

E

Breaks I'm gonna start the stream face.