Numenta Numenta Journal Club, 12 Jul 2019

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Neuromorphic Computing Paper review with Priyadarshini Panda

Description

We have a visitor who recently finished her PhD at Purdue, and will be starting as a professor at Yale in August. She has an upcoming paper to be published in Nature called "Towards Spike-based Machine Intelligence with Neuromorphic Computing". She will be discussing this work with us at our Numenta Research Meeting.

A

In certain states, okay, manic.

B

C

Hello, everybody we're gonna, get started real, quick here. Let me just just arranging these screens.

C

We will be I'll, be displaying our the meeting momentarily hello, Charles little Mark, Brown, hello, Falco, hello, everybody.

C

Hello, Daniel everyone's just getting their screens set up and then I will display them.

C

Looks pretty good.

C

It's a great idea.

C

Okay, I'm gonna go ahead and start streaming. You guys you ready! Oh okay,.

C

Yes, yeah, that's great, that's great! So I can swap switching between I'm really easy.

C

So I got almost ready, almost ready.

C

Today, I was working on this last livestream I work on this. Every livestream I'm, just waiting for them to say, go they're getting their screen setup I'm, not in control of the meeting.

C

Okay, just say when you guys.

C

B

C

C

I'm going to be switching back and forth between screens here.

B

A

I'm doing or just yeah or reopen it either. Oh look.

C

Guys, no, this is live so things happen.

D

To me this morning, yeah this.

A

Is what show so really happy to have? We don't actually need them under here. You see the professor.

D

I'm going to be a professor.

A

Ideal- and this is a review article- that he wrote a co-author on neuromorphic computing- that's not yet how it will appear, so we can't share the link yet, but what's it out, then we can share the link to the nature pack. So this will be on nature. Yes,.

D

I can like change a little because it has.

E

Been invited since.

D

Ordered meaning it was invited, and then it went to the review process. I.

E

Believe it was, if you just a minute, it was, except that they would've already modified the talk about that.

D

D

D

D

D

E

There was always, and today, computers.

D

Are everywhere, and there is technology.

D

B

D

D

We start comparing it to the brain now, of course, to the artificial. Very much you know has fueled everything and we have state-of-the-art results on several. You know cognitive applications right when you are looking at perception tasks that we are looking at active planning tasks, so artificial intelligence is very much there. So if you look back onto the basic ideas of the brain- and you know and neuroscience, computational neuroscience has had several interesting research efforts done to understand what the brain does, but still in that aspect, brain remains.

D

Was he an explorer so in this figure on the Left, what we tried to push out was what are the key foundational observations you know of all nuisance, research effort that has been done and what are the key foundational observations that can be attributed towards the remarkable cognitive capability that the brain has. So these, what we are essentially found out was one is the vast connectivity in the brain.

D

So you know there is a whole host of neurons that are connected to a host of synapses and the connectivity is on the order of one to ten thousand. So it is connected to ten thousand synapses. Okay. First.

B

Is this vast connectivity and it's not only.

D

Vast in our dimension and is a 3d connectivity, and why I am saying this 3d connectivity is something when you come to hardware, computer fabrics, building.

B

3D circuits are very difficult. It's all.

D

Laid out in a 2d manner, and that is itself current limitation from a hardware perspective. So one thing to notice here is that you know the distributive.

B

Computing that.

E

Occurs in the brain is not.

D

Just to redistributed computing, it's 3d, okay! Now! The second thing is, you know the brain is built.

B

D

Of basic primitives, which is in access and neurons less than access as your memory elements and neurons act as your computing elements right, there are other things I mean I like here on. You guys, happy panic. This.

B

Day, life as a rabbit.

E

Absolutely, but via.

B

Spotlight, you know.

D

Just to making simple dendrites.

A

Are part of the neuron, so you.

D

Can see the neuron is a maybe.

A

You could say it's a more complex.

D

So but then you know everything in the biggest build.

B

Out of that basic primitive now.

D

The next thing is that if you look at the combination, part of the visual cortex.

D

E

D

D

Now the other thing that is very important in the brain that we think is kind of enabling this energy efficient computing capability is the.

B

Temporal processing ability.

D

Now, what do I mean by the temporal processing ability is that information in the brain is pink student computed in the form of sparse events right so the dispersity in the bay, the neurons are communicating and there is normally start setting overall activity and overall, the way the gates are there, but the sparsity information exchanges also what makes up green that efficient and by and you know, if, like very recently, we conducted an analysis. I just.

E

Before that, I just won't say that drain on an average I.

D

Think people would know that you know clean on an average whatever.

B

D

B

D

Is in the brain is on the order of 20 watts right today, whether you look at a computer, doing stand and put a computer image recognition tasks.

D

E

E

That the brain is processing.

E

And you're, referring to specifically the information efficiency of music spikes and the fact that spiking, which is not the same thing.

D

So I think we have not run without it in the picture as we mentioned, and that is something that we were moving forward. I mean.

E

B

D

Since that lucky one.

E

D

E

You're also focusing on efficiency or.

D

The other ones they're really.

E

Not efficiency issues your mother to be categorical.

E

An occasion for spiking, as opposed to necessary the fundamental aspect of the cognition.

D

I think what you want to highlight.

E

D

Words of all these three activities, probably the temporal aspect of processing through sparks even human communication, as well as the causality which drove the brain, in course- and you know the causal way of encoding and decoding information. That is what is probably giving you that efficient aspect. The.

A

B

D

Are sort of the enablers on top of that, but those are organizations.

E

But from an energy.

B

Aspect, I think it's what.

D

We wanted to push out was it's probably this temporal exchange of information whose past events, that is the mean that is key to efficiency? Thank you.

D

Yeah I just want to highlight this part here, so you're actually I just know that you that in general today.

D

And we ran a simple example that people are embedded smart class processor.

B

And we're on a deep lucky.

D

Network on an embedded, smart class processor and this since it's an embedded device, it has practical scales right, so it will work on some limited factory about usability side. So this is a typical 2.1. What our factory and.

B

Using it without any my rough so that.

D

I things of the order 25 minutes so.

B

D

Be typically not what app will allow.

B

D

Would want right.

A

And I've shown you know a whole bunch of people who are trying to apply their stuff.

D

And embedded this issue.

D

So energy efficiency Zemke who go towards any artificial intelligence applications in a more practical matter: okay, okay.

E

Now, when I talk.

D

About the fundamental compute aspects of the brain, let's look at it exactly like today. If you look at the hardware, compute fabulous for the hardware ecosystem, they.

B

D

B

D

Know foundational principles which are similar to what the guillotine has, for example, even this every computer system today is based upon silicon transistors right. So it's a mass transistor, which is a metal oxide. So it's you know a standard. Mosfet is basically the computing unit, which is enabling every kind of technology of this. So there is a MOSFET.

B

D

Now this MOSFET can be organized into different circuit said you can perform any operation, whether it is an and operation or operation- and you know an amalgamation of all of these will give you some kind of multiply operation. Any corporation is a fair cetera, so processing, so you eat this. Mos transistors put it in an interesting way and image surface. Now these.

B

D

Now have to be put into a processor right which will do your standard computing. Now, when computing is done by the processor which can be held as like, you know the neuronal units, then there is also memory which will keep track of what exactly I need to either store, which will fetch, which will be fashioned which will be decoded when you're performing any of the operation.

D

So if you look at it in your standard computer break the main two parts, you have a memory unit and you have a processing unit right and the interaction between these two is more builds an entire system. Okay, now the main issue- and if you see that, if you just compare the male in the gradient memory and computer co-located right, I mean it's all intertwined with each other. But in this kind of a case in Maine, standard, bangla and processing I mean all everything. Today is a one time and computing engine where.

E

Your memory and your processor has.

D

Haptic, so what do I need, my that is just.

D

In general, every digital circuit, it has memory, you have the processing unit or CPU, and they communicate through a bus right. So.

B

Your memory, your stores, information, CPU.

D

Processes, information and.

E

They will be always some communicating.

D

With each other, so this is constrain right depending upon the bandwidth of the bus. The amount of data even push into the memory or you and pushing through the processor is going to be constrained. So this distance to this is known as basically, the constraint provided by it hierarchy will yield separating the memory and the processing unit is known as the memory vault water line, and this is what he leads to inefficient trust innovativeness. If you look at neuronal.

B

Operations, new operations are essentially, if.

D

One you have dates, which is memory elements and the nominal computational units, which is you know the.

E

D

Units as close to each other right, UVs are essentially feeding into or even on, a.

B

D

B

That somehow you want that some memory.

D

And the processing unit has to be together such that you can do the efficient processing and.

A

D

See POC gusta everything is one of the difference between a CPU and a GPU. Is that you see you, you have a sin even few or sing, or you.

B

D

Or a few simple course, but.

B

They can really process.

A

D

Complex logic, operation are straight, but.

A

B

D

What you have is right now, if you say that okay, you know what I not go for any complex logic, have multiple of simple course Hey. Because of the pattern. You have multiple of these simple course. Now you can analyze, it is so that is what we use to analyze your operations and.

D

B

E

D

B

D

E

E

Terms of the fact that you, these costly, active, CPU and.

D

B

Good thing to do.

E

D

But you see that you know what.

D

But this is what I want to say that you have to in.

B

Order to a minute.

D

B

D

Is the standard you know you.

D

E

Need to go from where traditional one I'm.

D

Going to work, you know that non modern architecture, I love going to the distributed computing, a framework or doing some yemeni-american. So that's the key idea: okay, so I'm just trying out so with the.

B

D

I was just trying to say that deep learning is also organized in hierarchy and deep learning networks. The hierarchical organization is worth less than learn. These different feature representation of the input, and that has become also heat. So you know you have to at this point of time in order to enable large data set recognition, problems I think one going the convolutional router on the decoder out is probably you know the state.

B

Of the art, if you can move it in.

D

Our much more simplistic manner with lesser number of players is the question, because that's what we would want you know going forward. However, this is the state of the earth. Okay.

D

Now, where does the spiking neural networks fit right? So what we said was electron arm. You.

B

D

In computing perspective, the entire is field of spiking neural networks, on your hope you take. It was diet by this. You know sparse, even given in our information exchange that occurs in the brain, so.

B

Spiking neural network so and if.

C

You go towards a.

D

Little bit of history of when how did computing coming to be covered mean, who proposed the field essentially wanted to emulate biological neural routines with transistors? So if you take a transistor, which is the standard or the other silicon limited, you take an ancestor, this transistor operates in different regimes like a transistor. Is nothing but a switch. They did. It is higher on stick and it has an off state. But if you, you know with certain you know, depending on in which regime you operated, he can act as a switch. It can have.

D

B

A nonlinear switch.

D

A

D

Cannot, with a nonlinear switch.

D

E

D

Some sort of to go to n conquers can be efficiently implement artificial intelligence, algorithms and for us at least the reading force. We thought.

B

D

Know because we want to implement efficient efficient if we make the network sparse, using even communication using event-driven activity, and then in that case you know, energy efficiency will be obtained because right now, unlike standard artificial neural network, stick each and every time any input comes. This neuron.

B

D

Be active right, however, if input is not coming, then this muon remains inactive. So.

B

Which is what we want like the more the sparsity.

D

In the activity, the.

B

D

Efficiency can be harnessed so.

B

This is the reason why.

D

We looked at the spy based information processing- okay, now I this one. Actually is this figure yeah, so I just want to go through this tiny, tiny figure which will really help us see. You know a kind of speed work. Other major you.

B

Know, discoveries and the advances that we made from.

D

D

Machine learning, intelligence and they never let hardware. This is what do you know the main.

B

D

Was fan me and.

B

Ever intelligence.

D

Right can we have algorithms that can do these specification or perceptive tasks now with new, robotic amputations, it was bootstrap. On top of this.

E

D

Idea was I think it makes sure that the hardware mattress algorithm as well as the algorithm can you know that's.

B

D

B

It's a very you know: existing logistic.

D

At work equally weighted between I'll go to the one hundred another.

B

D

Intelligence, so that's how many guys, how are you say, that's the main difference between what we don't normally I mean you look at both hard and got a minute right. You see, where are the mismatches well of.

E

D

Matches I try to go forward with it. Okay now here, if you look at the bottom figure.

B

D

If you like, what are the main innovations that we're done in hardware, this.

E

Was not necessarily motivated towards.

D

Artificial intelligence.

B

So they were motivated towards New Republic of beauty.

D

B

At that point,.

D

Of M, as I said, you know, no be computing, wasn't confessing just immediately biological neural routines later on it, artificial intelligence came into the be, and they do not for the same under the blue parts. Here they're all these innovations are motivated towards some new efficiency. So the fact that distributed memory processing.

D

You know and the blue curve and I said these are all motivated compute, how an ever better compute right so.

E

One moment build.

D

The first computer, then in the University of Pennsylvania, then they mention of transistor I, felt a switch did.

B

All the computing revolution- this is the 80s, that's a few already any just stop. At thirty years of the invention of the transistor people started thinking.

D

Of telling use the transistor to build the synapse any use a transistor to build, who have neuronal dynamics simulated. So.

E

They started using, as I said, the sub threshold rating of the transistor new.

D

Thing of protein, a transistor as a transition, as it is pretty.

B

D

And that's when Carver Mead was one of the fathers of this entire proposal and.

D

The fact that that would happen peoples are already.

E

D

Working on other elements that can possibly so besides using this silicon transistors, even when I started looking at different new materials, that probably do better, you know, computer so to the activity percent moving administers. So memory cells are essentially passive components which can act as memory elements. So it is an analog way of you know. Instead of having a bunch of transistors together, which are ready to use Sam's.

E

Rd lamps, that memory is stored and they are all composed.

D

Of you know a bunch of transistors, so mine is a basically an Euler. Computer components vary depending upon the state or depending upon how much of are you pushin. The.

D

Resistance changes and as you notice, that can also act as a memory element, and you will see today. People are trying to push if you can utilize memristors in order to get that enormous. Okay,.

B

D

As if I mentioned that, even since nineteen seventies I think that's when Woods traditional transistor scaling happened, and you know that's when it L got all the money started.

E

Making a lot of money.

D

Okay in the 1990s GPUs are formally introduced, so in 80s and 90s we were mostly working on CPUs, where he had. You know single core processing, but then GPUs were introduced by this, that you know for certain applications. I do not need these complex, single core processing. Maybe if I make my course very simple where they don't improve the critical need to simple operations, but I can then multiply that and I can do parallel processing and that way I will have much more efficiency. I.

A

Realize it was that early.

B

D

Now the second.

D

Spy based communication- and this is something very, very important and.

B

E

D

B

Should do sing.

E

Or whether you should do.

D

In synchronous processing give me no trouble, we have to bring all circuits at the right time right so.

E

D

Do not have any ports, because the way you will prompt the circuits that will determine how many performs the operation, but.

B

In the synchronous processing, we don't read about any.

D

Meaning, as soon as a unit is present, you communicated forever, which is efficient.

B

D

You don't need this extra walking mechanism and you need to do need to move it, so this is very efficient. It is not.

E

Controlling as.

D

E

E

Now we currently underline is encoding a range speed of things, so it's a essential component of processing yeah and why means were more flexibility in the hardware, because you discover these.

B

New components.

E

Of brain functioning, intelligent, accommodate.

B

D

I think that, like you, want to like I think this is the baby out probably were synchronicity. Is they kind of do a little clock but I think what you're saying is that in the brain are synchronizes.

D

E

D

Global channels, you need a Google clock.

E

D

E

Talk about those things in this hybrid secrets,.

B

E

In the brain we just gotta kick that I.

D

B

I mean this is not to.

D

Say that Jane is doing that I think the way to go forward with hardware are tapping into all efficiency benefits would be if you want to implement a spiking neural network in the best you know distributed manner, so I think having this hybrid global, Hydra synchronous, asynchronous no local through the channel of communication and the Communication I mean you know, he's component to sending some information to the other thing. Okay, now, if.

B

You look at the rest of it opposed 2000 and that's when.

E

B

Don't have to computing.

D

Started in a little bit of a shape, you see that some of the new chips started getting added up, you see the spinnaker chip and you have the new red chip. So these are.

E

All different I.

D

Mean these are eigen based upon digital substrates,.

A

All analog substrates.

D

A

Especially exploited analog.

D

Computation such that you know it can be in.

D

Stanford in about feeding.

B

D

Isn't Berkeley.

A

Wind, a nice start.

E

That's one: that's been selected here, the conferences just focus on me. You know like I, would you know? I spoke at ICML one year and I Miska I hate this guy, which is a whole section on it. But it's not every year.

D

Okay, so you know so the University Saturday new shape and that sort of thing. In fact, I just wrote that you know Blue Gene supercomputer, which was from IBM.

C

D

Accumulated activity of forecasting so.

C

E

Was a whole lot of interest in triangle.

D

D

If you look at the energy results, then on our supercomputer consume the order of something like.

D

So he knows about the travel needs.

B

D

They have done a lot of work, what better ones so.

B

Anyway, so what we're gonna show with it, if you look at the energy number, is that.

D

E

D

Know the different cooling system as a track for them so.

E

It was on the order not making in a few megawatts.

D

E

What we want to represent it.

D

Will be simulating five seconds of the cash, but.

B

The track is in.

E

Operating a twenty.

E

Yeah we speak the top, so.

D

B

So the energy gap was.

D

Still quite evident during that time then I did then the post.

B

2010 video.

D

Next, when we started looking at literally computing or building these anamorphic chips in order to you know, look at.

A

D

Right now, for the last few years, the effort has been can't even spike in deep networks right and in extreme convolutional hierarchy. Let's say an image net recognition or C for education, but then we have to go beyond these vision. Tasks we have to reinforcement. Learning is completely. You know unexplored in the SMN domain. I think that very recently in the year, 2019 some base reinforcements.

D

C

B

Fact is what he gets from the lack of appropriate learning: algorithms.

D

Is what is the you know, the bottom line at this point of lying and making sure that these chips are even doing what it should do? Energy efficiency perspective, this juice will perform better, no matter what has been another.

E

Overcome the other yeah.

D

A

Gonna build a custom chip.

E

B

One thing to say: I'm.

E

A researcher I'm focusing on these really, you know your history, it's like you know, I could study quantum computing and they have over 30 years. That's fine, that's good, but some of these are pretty big at first a lot of money and they hand out. Yes, I think they were surprised.

D

E

D

Is not to me right now what people are trying to have in the active research area is trying to see if we can, uh as I said earlier, like is silicon the best way to any.

B

Application or.

D

E

We go for our, you know our.

D

More beyond silicon technology and can be bringing going from our or can we come to this? You know a slight bit of analog processing, because that might also help because you have to be kind, Universal functionalities.

D

So my you know it's a mix of boolean as well as analog processing, and we are trying to look at you know: can we do in memory computing or near memory.

B

D

So, instead of having processor and separate you.

B

Know bringing them together or distributing them.

D

With a little bit of memory and compute co-located now the second thing we are looking at is: can we have our than Hardware co.design? Because right now you know whatever algorithms you build? You cannot have like this two separate, of course, algorithm is going in one direction and hardware is going in that direction.

B

With you know,.

D

Computing because energy efficiency is the keyboard, so can we build our algorithms to match to have the functionality or.

B

E

What we are trying to say no.

D

I think we are, what we think is is Hardware today has an auto oiler abilities right in any kind of processing, whether you do any in-memory processing or anything there.

E

D

The resilience these you know the arrow resilience operations operations and try to some of the abilities do that and if you have any processing or some of the emerging devices in their offerings of intrinsic functionality, he knew it that and pushed algorithm. So.

B

It's a boat, cider, dashboard and.

D

You want to see where exactly movie.

E

Gives mismatches years dramatically so that seems to ensure that rate of change slows down even to the harder start, cycling faster or they would start slightly slower. There seems to be an inherent problem of that.

D

If you look at what is it so for dramatic change in hardware, it's a long sellers and, of course, we know from research for a point of view when you look at emerging devices, so that still research right I mean you don't know, but the fantasy has excelled. L burning will progress much more faster and get? Hopefully, thinking that you know I got in progress is faster and today, with spiking neural networks. That is what is the key challenge, then mainly at least to the current hardware that we have. If you can address.

D

Some of the efficiency needs that the.

E

A

D

A

Hardware is not available for like machine learning. Researchers right if I want to run something on a GPU I can go to any.

E

Well, I, for example, I can't ship.

D

So all of these chips, you have to come up with.

D

That processing that you require, however,.

E

D

Know I mean I think we should here and work on it. There are some interesting functionality that is.

D

Okay, so if you look at my right in 1990 it was wolf and man who started looking at using spiking neurons as computational units in a neural network. So till then your network was there, but you will use an Armenian functionality like you know, Sigma. It was good learn once we started looking at you know, can we use this pi functionality and orange.

B

Highlights your these were.

D

All the algorithms, the.

C

B

Role of neuronal burdens, different.

D

Routine, it was understanding the range so.

B

That one of my upon.

D

The high early 1990s they showed that knowing them in the hippocampus that.

E

Is this form of spike timing-dependent.

D

Plasticity that occurs, which you know, plays an instrumental role in learning under the vagus in access to encode memory and then so so. Those.

A

Ideas were still there. If you look at the blue curve.

D

Still, you know.

A

The New Democracy I.

E

Mean is progressing at a much faster rate. Suddenly, you know.

D

You read books, then you know in 2012 election it was released and you know from there on deepening revolution happen. Fighting. On the other hand, if you see the main proposals like after the proposal of spiking Universe computational units, you know the only thing we would do was like trying to see if you can get a gradient descent version.

D

Of course I mean you know, thank you, and that was very difficult, so people started looking at you know spite prop, which is a rules that you know it is the using spike trains in order to descend, then.

E

People started looking at okay, you know what, let's not.

D

D

We try to see that whether.

B

We can take.

D

B

Neuroscience principles.

E

B

Doing better so.

B

D

D

Think that, at this point of time it's more.

B

D

Ie is being the burning, has its own gains and all we need to go to the next level of intelligence, probably.

B

It's better to take.

D

A step back and look what at what the brain, malice and try to integrate those principles said.

E

Or rather see what are the extractions.

D

That you can apply so I think it's a good confusion.

E

In this video time for the field.

D

Because itself IDs, we need to structure it a little bit a.

E

D

From a synchronous point of view, efficient communication still happens, but every unit is communicating at every point of time. Then.

E

Communicate still transmit some five seven spikes over some period of time and during that time I'm doing my.

E

So there's going to be the constant series of spikes, but I would also just ain't one of them, just single that that says. Here's a three bit value.

E

A

The issue is they: don't have populations varsity if they did? What you just said, every girl is there something.

E

A

E

A

Today's might be networks, don't have populations varsity.

E

Because it doesn't exist continuously exist time, so I can get my time.

A

D

I think I mean advantage of the.

D

Point zero one point: zero one and the inner Spidey encoded it probably will be encoded in. Despite of more spikes, let's say, point zero, one point: five one! That's it so, given the way you coded using decoding, given thanks it's as if this has this has 1 and 1 0 1. You know.

E

Quantizer context for job and if we quantize your activity.

B

E

Therefore, you know.

A

You know in order to get 8 bits of resolution integrated over certain amount of time to.

D

Get all the spikes.

A

There and you can look at the energy.

D

Suddenly, generally, you have to do a summation on.

D

D

As you said, depending on.

B

A good representation.

B

E

Spiking is an inherently necessary solution to any functional.

E

D

Right now, binaries networks are sort of synchronous form of these even.

B

Suddenly you can see but everybody's activity, zero or one motion, so these are spiking.

D

Network hardware: you will be able to energy efficiency so.

E

Why you do not go for a.

D

E

A different question: again: that's what we see that you know a binarized in and only thing is that in a by the resident.

D

You have to have it's not it's.

E

This in the distributed memory architecture- yes, very often because of the point on the 2d matrix. Just you also not relying bus structures throughout each individual spike, as opposed to providing a temporary packet with.

D

E

B

E

Try to create this.

D

So after that's right been be an inch FBI used some of these Network on checkout/return. Like you know, the dynamic routers process, information.

D

B

Soon, as you communication between the global having.

D

Said that he's at least the normal features that we have and it's true, not leaning on spinnaker, they all utilize these aspects, whereas GPU, which is our like a standard CPU. They don't have that kind of an architecture right. It's standard there.

B

Is a hold out of memory and of course we bring in these.

D

Hierarchy of memory cache etc. So you try.

B

To bring some of the memory within the CPU, but it's still not this.

D

Is more efficient, Minicozzi.

A

A

Think of our systems very sparse population battle sparsity. You.

E

A

Know course, coding- or you know, different types of encoding where scalar values are population level, not in terms of a rate and but you most of the stuff is zero. Wondering would.

B

That kind of a trade-off which I think is close.

A

To go the way it works.

D

You're, not exactly thinking in that cause our dependent this nothing for temperate. If.

A

D

A

Out of everything's course coded, and you can.

A

D

So this is and.

B

I just wanna highlight.

D

The current active research areas from the you know is why we want to use neural networks to understand the brain mechanisms. Then how extends networks to state of that is? We want to see if we can wanna go beyond vision. Go from. You know, title to corporate reinforcement, learning or continual learning tasks within a cycle neural network utilizing this extract and pure dimension. You have in order to you, know, address some of these aspects of.

D

Secondly, envy, and one thing is that we are using great coding mechanisms, because today we going to have data, sets that give you spike the spiking dig assess everything. Is your image right. So that's why we're using any sort of coding mechanism to convert an image data into spike data? But if you.

E

Had a neuromorphic guess.

D

That lets say that had a gig that could monitor.

B

D

Changes in the environmental activity and it will be winning back exactly so there. The temporal relation is very much intact in that dataset and using that data set in order to do image, processing and doing learning if it works. That would be even more interesting and that could be very, very efficient right. What.

A

About speakers yes,.

D

Videos buttons with videos- it's all- are still pixelated frame data right, so we have to come up with a way in order to extract out the rice bike signatures from those video such that you can. You know if you let's take in a video, some of it is moving. You need to expect out those object. Movement related insulin.

A

D

Dominant flames is not so yes, as I said lekha. Is that the friction of that it will give you just.

E

The standard it.

B

Will be a blackened body image right, so you you.

D

Got an information processing given that you have a good.

D

D

And that's why immediately she's not still a deep learning, they use an optical visual flow, but also they have the standard Excel data going in such that you know get.

B

Out get that relevant information. How.

D

Much time out so having a one-day.

B

Event but the most.

E

Important challenge today is how.

D

To develop good training, algorithms for a service, and in fact, so the problem is that, while these narrow fit ships are present, we cannot use them to develop training and go to their site. It is a restraining. You want to go towards the back, face periodicity and give you the good at doing that risky. As soon as you reduce the back size to 1, you will see that the GPU power conversion is pretty high and the GPU is not I mean it completely.

E

Render the whole of.

D

A GPU useless, so you want to, if you have these three moves to a fight or cafe or tensorflow, and if you want to add a proper spying, you know Network algorithm. Can we utilize? These already available frameworks are already available functionalities for corporations or you know any normal operations.

D

A

See that in the near future be able to and my pipe article in.

D

This interface.

A

A

B

So you make it a good automatic differentiation.

D

D

D

A

D

So now it's going to what exactly is: what are the intrinsic functionalities of exciting Network? We look at okay and I, say, like us, I think uses a different, neuron or function.

D

D

It's very simple: this.

A

Is the leaky part the fact that the membrane potential the case.

D

D

You have the motor neurons, which are against fight neurons, because you know any spiking Ron will August interact with spikes, so you have the input neurons, which are denoted as three neurons and then the would you not sit in speeding it if you can hear its information for some amount of weights into the post a blonde and this is which is the output so.

B

E

Of the neuron has this.

D

In an interesting, are we integrate fire dynamics in order to model what kind of produce? So you are you doing the same thing that the equal spike get multiplied with the waves and get stumped before you push it through the output? Now the nicotine is a functionality which is denoted by that signal. Be there it's nothing, but the membrane potential which I will, which is known as the activation function activation function, which is the blue line there. So the membrane potential keeps on increasing as whenever arise by input side. Is there.

D

B

Process, that's.

D

Why we produced this insight, which is the output file, so the integrated functionality, because responding functionality is finally and kompis in the entire neuronal activation? Okay?

D

So so- and this is how you know the neuronal, many others and, interestingly enough, you see the scheduling. Functionality is what accounts for that discontinuity in the neural functionality. So you cannot exactly use standard, gradient descent mechanisms, and you know you can expect if there are expect that it will go well, because that is a normally is a discontinuity there, which will be difficult to differentiate in your standard, really decent mechanisms, but pre.

B

There is a lot of work trying to come up with the proximity.

D

Dataset mechanisms which have been talked about a little the next now.

B

D

Equation to model your lis, you not in general now now. One thing that I was fighting very excited about is that you can actually use the relation between the output and input spike in order to conduct this, and you know inspired by a possible which is independent custody. So.

B

We have extracted, despite having dependent plasticity in a very exponential learning ruler.

D

Shul and depending upon, if the presynaptic and postsynaptic spikes, if they are close together, then there will be a higher correlation and the wake of rate will be higher. So speaker.

E

Is not of hebbian.

D

Plasticity so all.

E

New loans that fire together.

D

Should buy back at the rate, so that means if the correlation.

B

D

The pre and, of course, fighting is very, very very strong. That means you will.

B

Have a larger bit of the correlation is B.

D

Then you will have a smaller weight of it, and would you in the correlation, then the post fires after that lead right. So if, if the causal effect is present, where the post fires up the place, then it will be a positive 800 and if the causal, if it is absent, it means there's a depolarization. When the post fires before the three, then you will have a negative area. So this.

B

Is the standard thing.

D

And we have shown that we can do interesting, that initial applications with that now.

B

One thing I want to show you is that till now, even in SNL's.

D

Get trying to use your standard benchmark data such like c4 image, menorah cetera in order to conduct study right, but.

E

If you don't use their original image as.

B

D

Because you have to feed in spider insertion so.

B

What we do is we do rate for it.

D

So, with real coding continue with that, every.

B

Pixel value, none is.

D

B

As a spike train now, what I'm, showing you in this case.

D

Is let us say we take a 15-time step spike train and we can pour this original image pixel information using poison folding, so, based upon the Poisson distribution, we will uh I will put some spice so I equal.

B

D

See that you know the.

B

Image is original rate and.

D

Equal to find what I'm, showing you here is.

B

The summation of all the.

D

Inputs always fights in events that you could have generated between equal to 1 and equal 5. So if.

E

You sum all the maps together, you start.

D

Seeing their peers of the information sets coming back and.

E

Equal to 10 again the summation rate.

D

For expect unison, so.

B

What did I say that is possibly an ordinary.

D

B

A

A trade-off there willing to increase the.

D

98, which is the original, but.

B

D

You said if the network is trained well forcing maybe we can't stop influencing and equal to answer. You.

A

Can offer for the wrong man for the software you can infer that.

D

D

D

So the question is, can be.

D

D

D

And this is more stuff. Some of the main things have also come out from professional photography for you. So.

E

A

D

It is completely unsupervised which is good for us. It is local learning, so you, if you can think about it. You know with us from a perspective.

B

It has a lot of advantages. The.

D

Problem is that it gives you limited accuracy. Okay and even it is only like, as I showed you are shown that does not work for multiple years. You cannot go deep with a local dining room, I mean you can probably go team, then you have to be everybody's.

B

But then somehow.

E

B

D

An and back propagation in order to find punic, so the problem.

B

D

You scale it up. That is the main problem with actually okay.

D

E

D

Want to talk about, then the cart like okay, you know. If actually we learn your scientist learning is not working. Can we actually take advantage of the standard range training and about.

B

A QRS Ellen, so this is what conversion is so.

D

You need to create in, and then you can borrow, it was hiding functionality hiding your network, and why do we use this conversion and what do we achieve? Mystic Roshan is essentially taking that lady of the Dell. You do not take the subway. You want to exactly match it to the constraints of a spy.

E

In your rod, so given.

D

That you know a spike in Iran will have a leaf and constant will have some refractory period or will have its Yom Hashoah, which it will spy. You have to.

E

Adjust all of these values in the basis.

D

That the rate of the artificial radio would do not matches to the rate of the I mean you're on given the every dimension. So that will give that motion mechanism and in.

E

This is the interesting thing.

D

Is that this is the movie, this is the state of that we have been able to convert imaginary.

D

Neural networks, but you know one or two percent of accuracy lost, so this is giving you the.

D

Network all of the computer advantage is coming to being so. You know.

B

D

Is partially everything comes into being, and these.

B

Are, as I said, the.

D

Energy, stable is because of the fact that you are performing an active minute operation, not.

B

A multiplying a comedic operation.

D

So you still get that thinning advantage perspective, but.

A

The training is done.

D

So good thing is that review of these five cross train works with the burden of training in the sky. That is gone, but still you would eventually want to go to words so.

A

That could be a way to integrate like work, you train with your normal, my culture than you. Yes,.

A

B

Problem is not you is that the.

D

Dnc right I mean and the fact that you want to match the output of the ALU have.

E

To give it that much.

D

Of time, like you have a very high number of latency such that you can, you know integrated over a good amount of time to match down. So this is a huge concern, so.

B

Different to address that we, it is affect whether you can go.

D

From 2,000 extra 200.

E

D

He started including some approximate gradient. Descent is learning yields by a neural networks. So what is approximate.

E

So I go you what.

D

Exactly are we trying to do with approximate backdrop, but here what we saw was the prose was either able.

B

To get miraculously.

D

The problem is, we cannot scale it down, I mean you know, we could start today. The best accuracy they achieve goes on cPanel data set on a vgg. You know eight nine line network, but as soon as you put on signature on problem start happening and that.

B

E

Today we are doing.

D

Training is also because of my customers using GPUs, but, as you can imagine, either take use modified ashram of exactly accounts for the are even.

B

Driven operations.

D

So taking I mean you know, taking a large network training it on a large raid us and it has it it's even more time. So those are kind of the cons today, but Yahoo Yahoo I will see if we can use a drop. Some intelligent mechanisms in order to scale.

B

D

To beta readers and ecology and I want to just highlight.

E

D

Got star: okay, let's just.

D

B

Yeah highlighter is.

D

The stochastic spiking neural.

B

Networks, so what we're trying to say is that till now we know one of these ellipses.

D

Factor of a n, n or STD be the waves activations that are all on the sort of like is the base are especially floating-point values right, but can you go to his family values so that question right? So can you have finally finalized fields and in.

B

Fact we are trying.

D

To see if you can come up with the stochastic version of acidity googleberry, instead of using an exponential correlation base, we don't even have a standard.

B

You know stochastic weight, update and none, given that there is polarization again.

D

Increase by a fixed amount, so you know from our programming.

B

Perspective from a hundred perspective, it would happen with.

D

An N is exponential. This is again switched so an imaging.

B

Device technology that.

A

We are looking.

D

A

Has a lot of advantages.

D

A

Is this binary ways.

D

So you are good increase or decrease, it's just that when you decrease it, whether they should stay at stay. You know whether it should go to the street, which is one decrease, is determined by this point. So so there's.

A

Still some non binary quality.

D

And yeah, so this.

A

Is actually a notice that.

D

The probability of switching between current state and the next state, the.

B

Probability of the switching can be.

D

Acting in from that correlation.

A

Mechanism- okay, so that is.

D

An interesting way of doing binary based learning is now how much you are using correlation in order to determine switch to the next beer.

D

A

With that here bill together, 98.5 yeah.

D

So, having said that, so these are the main and personal training training site. Now, I just want to go to you about that approximately innocently late, having that's one of the major efforts that we're going to do in order to scale it a lot larger.

B

D

With lower intensity, so what exactly is the thing you're trying to achieve with the proximity? Radius in training? Is that let's say you have this neural network? You do standard forward, propagation and back propagation rate doing forward propagation. You will get a delegate activation values and you will calculate the error. This.

E

D

Then be propagated to calculator of these now for doing, ah update ways if you employ chain rule this quantity, which is the F of F dash, which is nothing but the difference. Now, the gradient of the activation unit that actually comes into play when you calculating any wakeup gives you the chain and if, in a case which I showed you earlier.

B

D

Discontinuous model.

E

So that it factory does not exist.

D

E

We are trying to see that can get on with.

D

The approximate version of F, so that's what many binaries you don't go through. They go through some approximation of you know using a stateful estimate or something so here, given that we have a temporal correlation. So can we come up with a move on in your way of approximating that function such that we are able to integrate all the temporal constraints while doing gradient at leisure? So all the efforts from land from our group across every group today who are doing our base, building design they're trying to come.

E

Up with a better approximation.

D

So this is the crux.

B

D

It and very recently, my drama only neither able to come up with this interest if I proximation remodel, which was able to give you the state of that acted as MC.

D

I actually just wanted to highlight.

B

D

The he Harbor perspective that we need to do that going forward and remember. This is.

B

Not just you know much.

D

Because I interview what.

B

Ever had but architecture that evident now we will talk.

D

About the distribution of memory, these can even absolutely employee networks. Okay, it's just that. I think that sparks event-driven communication. You can add another level of.

B

D

So you look at I want to just highlight two of the major new romantic ships that could develop. One is diluted and one is too hot and you.

E

B

Is an analog design, I.

D

Do not mix it with the Sun in fully. You can remember what is the difference here. You, your started, bringing memory and the do not computation together in this piece also this red units. They have this memory and RAM together. If you look at it, so what you are trying to do, is it bringing.

B

D

A separate memory, separate processor architecture- you have distributed it across so I love.

E

This one, this unit will do some computation. This we are introduced of combination. This will reproduce of computation. They will all communicate across each.

D

Other and that is more efficient when you look at it, be like large network which has competitions, these kind of activities has more efficient and and so going forward a distributed way of communication using this interesting form of.

E

Rowdy right because now you don't have a.

B

D

B

B

D

Is coming if we communication is happening which of the chip itself is.

B

In two units, this.

D

B

D

Aspect of this local at a local hardware, has direct correlation with super learning boots. So you know that that is not advantage again. Okay and I just want to say.

B

You that address.

D

Equilibrium and representation, which is very much important, even if.

A

You, if you think about my propagation, which is not a local learning rule, the information still flows in a localized.

B

A

Just that I guess you can't really do so much in parallel, but.

D

The problem is I, think about it. This way, I just let.

B

Me dizzy: 16 I hope you.

D

Know each of these you know, units that say and put two layers of the.

B

D

B

This is a fast.

D

Way, this is the second.

D

D

Tony, you know, generally, you want to communicate a point.

D

You have to say the address.

B

Should you just passing the.

D

B

If you look at the.

D

D

Information is maintained because there is the first by Thompson. So first you send the first packet.

D

Information about where this is.

A

Coming it's just coming as the order. Yes,.

D

Not necessarily the.

A

Latency between so.

D

When you have a deep.

A

D

Water cans have information about; it should.

B

I get or should.

D

I not, but generally you know it automatically, maintains the.

B

D

As well, you can see between the next eight hours so.

B

This is a very interesting.

D

Thing about the address, even from in the presentation and that's what I want you to highlight that why it is the spiking Network that receive a reservation. So this is one advantage that you.

B

Don't have to maintain a lot of.

D

B

D

Is latency? You also maintain the Bayon. You are etching spydra.

B

You know the weather, this one has smaller this one. How do you go that it.

E

Will address even the presentation it.

D

Makes things very much because.

E

A

Is kind of maintained given.

D

Them the encoder and decoder action.

A

Americano, so you can say they both came and exactly at the same time. So.

D

A

Probably so rare that.

D

So let's say things can come at the same 94, so you can.

D

Just one pin what you can do is you can have an extra another information in the package saying that it is coming with you can actually in then, when you the receiving chip in the decoder, you can actually have some extra surface of control in order to, but.

B

E

B

So I am just I am saying.

D

By addressing in the presentation is important because, from my spidey perspective and.

B

D

And it's been easy incoming, okay.

A

This is gonna, be done in the discreet system, exactly.

D

A

Does not require.

D

A

D

And so, as I said, okay right, if you want to enjoy the nonlinear functionality, let's say right: you want! So if you have a monitor functionality like this, you do not wander. You can utilize a transistor in a different regime in order to implement a nonlinear functionality. Okay- and this is an Universal Charlie which can be a sake model. Another study now.

B

E

Standard silicon transistor that cannot be implemented, given that you are operating.

D

In to that analog regime of this now there are a host of other devices that are showing you on the Left phase. Change memories are vans, stdin ramps, protocol transistors. These are all sort of these. Are you know, post silicon devices, post CMOS.

B

Devices that have a lot.

E

Of non-linearity embedded.

D

In the physical dynamics, so the ECM changes- oh I, sticky mm changes. Its state from you know, on state to off state it's a very nonlinear manner and you can utilize it in any kind of a needle.

D

So what I'm going to say that today, the physics of the devices are actually being used to emulate Universal active functionality and the interesting things as all these devices? Are they interesting physical.

B

D

Should even one to one match the developments in our functionality? Okay, but the main objective of our device is to act as a switch right either it has monster or it has now.

A

Let's see nameste.

D

You have one of these.

A

D

A PCM, our agra- these can have multiple steps of memory. So what my everyday.

D

D

D

D

A

D

A

This now that people think can be fabricated, so.

D

People have the physical Rockstar implementation and they're trying to be human, logical, crossbars, so so logically doing crossfire operations rather than physical. So this means you have a memory sticking at a cross point which will help you do or logic which will need to do an order of operation, but you can remove it. Take if I have a SM.

D

You will do interesting now.

E

You can transistors in order to.

D

Locate some of these issues so.

B

That is, but then progress have a lot of money. Still like one of.

D

Those you know.

D

B

New sativa image, which.

D

Has these programmable so.

B

That means that these.

D

Technical actors- okay, relevant because it has different states address away so.

B

These are given sitting at the cross points, which is show another thing.

D

Okay, so now, let's see the holes are the lines are the inputs and the vertical lines are the outputs? Okay, so you send in hang on I I agree, then the high one is the voltage that is going in and this is a resistive element. So from also we know that kind of a voltage passes through a resistor. It gives you some height thing so.

B

It will produce from connect here that is course.

D

B

Again, if you pass a voltage to.

D

This will pass the same thing will give you something now the Kirchhoff's.

A

B

D

There is current coming here to community current compañía Kirchhoff's law at the end of this particular.

B

D

You sense how much of corrective or sense it will be: I, 1, plus I, 2, plus I 3, and that.

E

Is nothing to this operation.

D

That you have created that Sigma bi into the Sigma.

E

Bi into WI is your standard.

D

Weighted input summation operation so that just utilizing this increasing architecture place in the power of the memory of anyone who have even you.

B

E

Get water on up.

D

In such an efficient.

E

D

So you don't have to have separate is that it's doing the fetch between different.

B

Lilliput arrested the memory and the computation are all together right.

D

B

D

Very interesting about the crossfire, and if you look at it from a political perspective, when I showed you this here, each of these units had that over multiplier cumulative memory, banks right it is like mu Rong's, where they ran, was they're separate, but I.

B

Don't need anything.

D

D

So the demons can be some.

A

D

But the best part is that the remedies will get you.

B

E

D

Unit so each of these leg.

E

D

As they can shown, as you see ya, these are essentially computing. The weighted, sufficient or.

B

Product operation, as well as in and out.

D

And then you have the initial.

B

Communicate so.

D

Right now now the disadvantage.

B

Which you will know this year is: let's.

D

Say if I have a sparse network right, so that means it is fashionable. I need only the first three things and probably the last it is I. Don't need this so.

B

Right now, the problem with the cross. Why? Because.

D

It's a physical way of computing everything, sparse vector multiplications are not exactly useful because you know the.

B

Cross that is already mean you have already had it into the copper, even though you specify those will.

D

D

So it is very easy: I mean for prosper, that's the problem. There are interesting days in which you can come up with algorithms in order to, let's say in order to make sure that the PA you specify your matrix, it still measure its.

D

B

D

So you just say you look a link between this cross sport in this cross point. The voltage difference between this crossword they'd be able to program the and.

A

We'll remember that there was a spike there. Yes,.

D

Status is that's so yummy. These are little. Crystals are basically memory.

B

D

B

The way they change the sense that is.

D

Memory inherited, so it is very interesting about these devices. It's.

B

Just that the problem with this device is that they're not as good.

D

A CMOS and C Mon is very controlled programming here it can go here and that's right. There's a lot of research I'm trying to see that if they even can replace CMOS in the future.

B

D

And the hardware, if you look at it in your company you're trying to see where do they match? Where is that right? You know animal cognition of night marriage where we can take advantage from only our things and so from this people, if you'll be in.

B

The final section we are trying to say what are the prospects with algorithm art, deco design.

D

So, for here is analog computing with a new belief in good words, dot product, like the crossbar of things.

D

But programming algorithm itself is a resilient may.

B

Be it will be, go.

D

Clearing, so that was what that's, what we are trying to say that you know with the proper.

E

D

Can probably address.

E

B

The hardware and.

D

Abilities- and this is one of them and you.

B

Know trying to see if you can you joke, like learning in.

D

B

D

So- and you know, stochasticity lots of unexpected has a lot of advantages from another perspective and perspective.

B

D

All tolerance of robustness: equally, you do have the proper.

D

What is he like? Oh this is giving the best result. No, even from my side, I mean should.

B

We go also I think we'll.

D

Know if you are in human perception, cows may be in the final classified. It does not make sense to have a spy base off max. There may be the standard soft marks. So you know approaches are more efficient for accuracy for energy efficiency, so all of that has to be taken in so with that this paper just tries to make a case when you don't want the computing, but we have to come up it's in a mystical reference from both.

B

That's great, thank you.

B

A

C

It takes everybody for myself.

C

All right, thank you for for joining the livestream. If you like this content, please please give it a thumbs up and subscribe to the channel. It helps. You know my bosses communicate to them that what I'm doing here this is useful, stay tuned for another live stream in less than an hour I'm going to be back online, doing a live coding, building, HTM systems live stream and I am taking off because I got to go, eat lunch first, so see you guys in less than an hour.