Intelligence and Machines: Creating Intelligent Machines by Modeling the Brain with Jeff Hawkins

Intelligence and Machines: Creating Intelligent Machines by Modeling the Brain with Jeff Hawkins


– [Narrator] This program is presented by University of California Television. Like what you learn? Visit our website or follow
us on Facebook and Twitter to keep up with the latest UCTV programs. Also, make sure to check out and subscribe to our YouTube original
channel, uctvPrime, available only on YouTube. (electronic music) – Welcome, everyone, good afternoon. My name is Bruno Olshausen
and today it’s my very great pleasure to introduce Jeff
Hawkins for his second lecture as Hitchcock professor. So in 2002, Jeff founded the
Redwood Neuroscience Institute, which would focus on developing
a theoretical framework for neocortical function. I eagerly signed on as one
of his first scientists. RNI grew into a intellectually
rich and vibrantly resourced institute, and Jeff
was the person who brought life to it and made it
an exciting place to be. He grew and shaped the institute
into a group of individuals sharing a common vision
to develop a theoretical framework for thalamocortical
function, and who worked interactively on a daily basis. We held the weekly seminar
series featuring distinguished neuroscientists from across the globe. Jeff was known to interrupt
speakers at the beginning of their talks as they were
just describing the problem they were working on to ask,
why would you want to do that? What is it going to tell you? This produced a range of
interesting reactions. Some perhaps have never
been asked that question so point blankly before. But almost always, it lead
to an interesting and thought provoking discussion, and these
seminars went on literally, for hours. The speakers loved it. We loved it. And everyone learned a
lot from these exchanges that were in large part
driven by Jeff’s intellectual curiosity, and his intense
desire to understand the science. In 2004, Jeff co-authored
together with Sandra Blasely the Book on Intelligence,
which lays out his ideas about the cortex as the
hierarchical membrane system that learns from the
environment and does prediction. This book has inspired countless
students to enter the field of neuroscience. I know this because my inbox
has been flooded with their emails since then on a daily basis. And also, I was on the
neuroscience admissions committee, and read many of their stories
in their application essays. These stories go on, these
stories go something like this: I thought I wanted to go into engineering. I thought I wanted to go into medicine. Or into law, or God forbid, banking. And then, I read Jeff’s book. And now I want to study
neuroscience and work on models of the neocortex. Some students ended up
interviewing for our program, and they arrived asking questions like, do you believe in a
common cortical algorithm? They got that question from Jeff. And, that’s a great question
for a student to be asking. In 2005, Jeff started
Numenta to push forward the development of intelligence
machines based on models of brain function, and
today he will tell us more about the models they have
developed and the principles that make them work. In that same year in 2005,
Jeff gifted RNI to UC Berkeley, where it now functions as the
Redwood Center for Theoretical Neuro Science, part of the
Helen Goals Neurosciences Institute. The Redwood Center
Endowment funds students in the Neuroscience Project
Program in addition to seminars, courses, and other programs
at the Redwood Center. The center now provides a
rich, intellectual environment for students doing work in
computational neuroscience, and their now 10 PhD
students who have graduated from Berkeley having
done their thesis work in the Redwood Center. So I would like to thank this
opportunity to thank you, Jeff, for what you have
brought to the Berkeley Campus not only in the intellectual
contribution, but also helping them to create an environment
that enables students to learn and to grow and to
contribute to an emerging area of neuroscience. So please help me in welcoming
Jeff Hawkins for a second lecture as Hitchcock Professors. (audience clapping) – Thank you, Bruno. That was a very nice introduction. I should tell you a story about Bruno. When I started the Redwood
Neuroscience Institute, my first question, who would
be crazy enough to come work here? I mean, it’s a new institute! You know, how do you open up? You put a shingle on the
street and hope neuroscientists walk in the door saying
hey, I saw your sign! And, it’s a very odd career move. You’re going to a non-affiliated
institute, and you know, it’s not leading to tenure
or anything like that. But, the people who really were passionate about understanding how the
brain works and the neocortex were willing to take that gamble. And Bruno was one of those guys. As he said, he signed right up. He says I want to do this. I care about this. Bruno became my right hand
man, or my other body, my doppler game or whatever
you want to call it. And, the great thing about
Bruno is, is he has this encyclopedic mind. He knows everything, everybody. It’s like he can memorize dates, and I can’t remember anything. So I just follow him around
and say Bruno, who wrote that paper about such and
such, and he goes oh that was so and so in 1964. Great, thank you. Bruno, who should we have to
speak about the basil ganglia? We should have these three people, great! Let’s do that. So, been great collaboration,
I’m happy to have the introduction, be introduced by Bruno. Yesterday, I have two talks. And these are the titles. Intelligence in the Brain, that
was the title of yesterday’s talk, and Intelligence and
Machines of the Brain in today’s talk overlap. And I will talk about that in a bit. But I just want to ask
something of you right now. I’d like to know who in the
audience was not here yesterday? So if you were not here
yesterday, please raise your hand. Okay, that’s a fair number. There is an overlap between
these talks, and I’m gonna, I have to decide how much
to repeat myself and so on, but I’ll try to manage that. Okay, just to remind you
for those that were here, I can’t imagine doing anything
else than studying the brain. To me, this is the most
interesting thing one can do. We are our brains. Everything we do with our brains. Our life is our brains. Our questions is the
product of our brains. Our knowledge is the
product of our brains. Our art, our science, our
literature is what we are. The ability to even
know anything of course, is a product of our brains. And so to know what humanity
is, you really should, I just wanted to know what am I? How do I do this? How did we come about? And where’s all this gonna go? So, that’s where my interest
in this field came from. Just deep curiosity about
myself and other humans. And then, as I got into it,
I realized that you could build machines. So, they work on those principles. So, I started out, let’s
see, I had little problems with this yesterday. Hopefully it won’t be a problem today. Okay, I started out
basically trying to answer these two questions. I said, okay, how can we
discover operating principles of the neocortex? And then I realized once we
understood those operating principles of the brain,
we could build machines that work on those principles. This is the order in
which I care about them, but they go hand in hand. So there you have it. Just to remind you, for those
of you that weren’t here yesterday, the neocortex
is about 60% of the volume of the human brain. It’s the big wrinkly thing on top. And it’s the location of
all high level thought. Language, science, high
level motor planning. Anything you can tell me about, is stored in your neocortex pretty much. And so, it’s the system we’re
interested in when we wanna deal with intelligence. So, this is the process again,
I showed this slide yesterday this is the process I go about my work. I start with anatomy and
physiology, this is the detailed, very detailed knowledge about
how brains are constructed. We use that as a set of
constraints on our theories. Then we develop a principle
for how which that works. And then finally we can
develop hardware and software solutions that implement these things to build intelligent machines. Yesterday’s talk was about
the first two of these items. And today’s talk I’m gonna
have a little bit of a review on the principles, and then
I’ll talk about the software and where this is going in the future. So here’s my outline for my talk today. I’m gonna do a brief history
of Machine Intelligence. Very brief, I’m not trying to be complete. It’s very opinionated, so
just take it for what it is. I’m gonna then tell you how I
define Machine Intelligence. What are we trying to do? How will we know if we
achieve it in some sense? And then I want to do a little
bit of a review about what we’ve learned about how
the neocortex works. And then I’m gonna talk about Grok. Grok is a product we’ve built. And it’s an illustrative
product because it’s based on these principles and it’ll
tell you what we can do today. And then I’m gonna talk
briefly about the future of Machine Intelligence,
and answer in the end, why should we do this, why do we care. I’m gonna start with our brief history. Now this man, Alan Turing,
very famous, he was a British mathematician, most of you
probably have heard of him. He was one of the guys
who actually helped us win World War II. He was a key in deciphering
the Enigma Machine, which was an encryption machine for the Germans. A real hero. But he was interested, he was a founder of computing science. And he wrote a paper starting
in 1935 and then continuing on from that, which is really
a big foundational principle about how computers work. Now, this quote is not from Alan Turing. This is my paraphrasing
about what his paper starting in 1935 was all about. He basically said you know,
computers are actually a universal machine,
they can model anything. If there’s anything you want in the world that can be modeled, computers can do it. There’s no other reason
to do anything else. And all computers are essentially
the same, so any computer can do this, they’re universal machines. And this is really a powerful idea. He then wrote another paper in 1950. Now he was very, very interested
in Machine Intelligence. He said oh, we can build
intelligent machines. And he wrote this paper in 1950. There is a picture of
it on the right there, the cover on it. It was called Computing
Machinery and Intelligence, and in here he basically
said you know what, we can build intelligent machines. But he was worried about something. He was afraid that people
were gonna get in endless arguments about whether
this could be done or not, whether you know is there a
soul in the machine and so on, he just didn’t want to get into all this. He said this. He wrote this in the paper. He said, I don’t want to get
into arguments with people. He said, let’s just agree on the founding. I’m gonna come up with a test
for machine intelligence, and if we pass this test,
then we’ll just have to agree, you know, proposing that we
can agree that the machine is intelligent. And the test that he
proposed is now what we call the Turing Test. And, the basic idea of the
Turing Test is shown in this little picture in the bottom here. It’s just you know, he
called it the imitation game which someone sits a tele-typer
terminal and types questions to a computer and a human. And then he looks at the
answers, or she looks at the answers, and trying to
decide which is the computer and which is the human. If he can’t decide, then we have to say the computer’s intelligent. This is clever, but a really
bad idea in my opinion. It set things up for a lot
of problems going forward. First of all as I say here
at the top of the slide, it kinda set the bar of
that machine intelligence is about the human behavior. Is like we’ll know when we
have machine intelligence when it’s, you can’t tell
the difference between an intelligent machine and a human. Well, there’s a lot of problems with that. First of all, you can imagine
very, very intelligent machines that would not
pass the Turing Test. That maybe speak another language. Or don’t speak a language
at all that are still very, very intelligent. We have lots of intelligence
in other species on this planet: dogs,
cats, monkeys, dolphins. None of them would pass the Turing Test. On the other hand, you can
imagine some of the machines that might pass the
Turing Test that really aren’t intelligent at all. So this set us up. But these two principles
together based that computers can do anything, including
model the mind, model the brain, plus our ideas to
replicate human behavior, was really the foundational
principles are what came of the Artificial Intelligence movement. And, this went on for many years. I’m gonna summarize the
Artificial Intelligence world in one slide. Don’t want to do anything
rude about that, there’s a lot of great work went into this field. Notice I call this non, no neuroscience. In essence, that’s my definition of AI. People trying to create
intelligent machine, but just don’t look at brains,
don’t care about brains, don’t think about neuroscience at all. On the left, I listed
just a partial list of all these different initiative
techniques people have had over the years. I don’t need to go into
them, they covered everything from robotics, there’s a
picture of Shakey the Robot, that was one of the first robots. And then a recent one, Azanog. We have the famous idea
computer playing chess against Casper Aug, we not have Google
driving cars which we can do pretty soon, that’s pretty cool. There’s a little picture of these blocks, see that was something
called Blocks World. It’s about natural language
processing, you ask questions about that. These are all things that
happened over the years. And recently we had IBM
and another computer playing Jeopardy. That was pretty cool. There also have been major AI initiative. The MIT AI Lab existed decades,
it’s still in existence, excuse me. Which was the center of
AI research in the world. There’s something called a 5th
Generation Computing Project. You might have heard of this. This is from Japan in 1982. And Japan announced that
they’re gonna build machines that are smart or smarter
than humans, indistinguishable by humans in all ways. This scared the bajesus out of
people in the United States. Says oh my gosh, this was the
time when Japan was really rising into technology. And people says Japan’s
gonna pass us in technology capabilities, and this got
a lot of people riled up. The next year in 1983, DARPA,
our military research group, they created the Strategic
Computing Initiative which went on for about 10 years. Billions of dollars were put
into these two initiatives. They failed miserably. We learned some things from them, but they failed miserably though. Then there was recently, you
might remember, the DARPA Grand Challenge. This the one to get the cars
to drive through the desert. That was very successful. So, there’s a whole bunch of
things that have been going on here for decades, and if
you tried to follow this, it’s very confusing. All these things, different ways. Here’s how I summarize these things. These are good solutions,
many of them were very good solutions, but they’re very task specific. They weren’t general purpose solutions. Google self driving car
can’t fold your laundry. And you know, it can’t cook dinner. It’s a very specific things. These were merely program solutions. Engineers design program
solutions, and with very limited learning, some of them
have learning capabilities, but generally, fairly
limited learning abilities. And, if you talk to AI
researchers, they will universally tell you one of the
biggest problems they had is knowledge representation. They couldn’t figure out, and
they still can’t figure out, how do you get knowledge about
the world into a computer. And you think, what does he
mean knowledge about the world? Like, well, what’s a car,
and what does a car do? And what are all the attributes of a car? All the different types
of cars, et. cetera. It’s a huge list of information
we kinda know, our brains know, but it’s very, very
difficult to get computers to do this. And this is probably one
of the primary problems of AI research has had over the years. Now we’re gonna switch gears
and talk about two ‘nother guys back in the early
part of the 20th Century. This is Warren McCulloch and Walter Pitts. They also wrote a very important
seminal paper back in 1943. Here’s a picture of it, had
a more, had a highfalutin title of A Logical
Calculus of Ideas Eminent in Nervous Activity. Not as simple as Alan Turing’s
title, but the basic idea in here is they said you know what? We can think about neurons
in brains like computers. Here’s what they said. They said okay, we know about
neurons, and we know that these are cells in the brain,
and we know they process information and they get these
inputs on their synapses. And they say, well let’s
model that with a simple sort of model. These are the guys that first
invented the idea of a model neuron, or an artificial neuron. This is the first time this appeared. And, this basic model they
outlined is one that’s still used all over the place today. And it says here, it says
well a neuron’s like a cell that has these inputs this. Well, these neurons have
weights on them of sum of them together and if they above a
certain threshold, the neuron fires and has an activation schedule. This is as best as they
could do at the time. And it was pretty good. We now know this was a very
insufficient model of the neuron this was very much unlike real neurons. I’ll talk a little bit more
about that in a second. But at the time it was the
best they could do, but it has some fundamental flaws. Then they went on to do
something more amazing, or interesting, or perhaps regrettable. They said well look, if we
design these neurons in a very, very weird way and put just
a few inputs of nothing neuron-like at all, we can
turn them into logic cases like ands, and ors, and nots. This is the basic language of computers. And then they can say why, guess what! You can make neuron-like
things, not very neuron-like, but we can call them artificial
neurons, and you can build entire computers with these things. You can do anything, and
therefore it’s a universal Turing Machine again. And, so they saying,
well, let’s look at that. And so people got this idea
like wow, let’s use these artificial neurons and try to
build you know, Intelligent Machines using these artificial neurons. And so that in some ways
was the genesis of the whole field of artificial neural networks. Which is an entire genre
of machine intelligence that’s been going on for decades. I call this minimal neuroscience
because it has this image to the neuron, but it’s not
a very realistic neuron, not close, and beyond that,
they almost completely ignore, most of these models, 90 some
percent of them, completely ignore the anatomy and
physiology of the brain and all the details we have
about how brains are actually built, they just built
these simple networks. And I’ve shown here some pictures
of some of the artificial neural networks you’d
see in the literature. They have lots of similar characteristics. There are some names of
different types you might see. There’s plenty of them: back
propagation, perceptrons was one of the first ones,
Bolzman machines, Hopfield networks, Kohonen networks. There’s two books that came
out in the mid-80’s called the Parallel Distributed
Processing Books, which created a whole flurry of activity in this field. And there was a PDP society and so on. That was going on. Here’s how I summarize them. The good thing about these is
that they’re learning systems. They essentially say we
need to learn from extremes of data, we need to learn
from data being presented to these things. They were useful if they’re
still used in many applications. But they’re very, very limited. They don’t do much. They’re essentially
classifiers of various sorts. You can give a pattern
that can sort of match it to something else. And again, they’re not
brain-like at all, really. So, no one, AI and Artificial
Neural Network are not even close to producing
something we would say wow, that’s an intelligent machine. We have, it’s just, a huge
gap between where these fields are and what most people would
think about when they think about intelligent machines. So, this lately has been
another activity in this area you might want to know about. This is, I’m only going
to give one example. This is an interest that
people try to do whole brain stimulation, they say wow if
you have all this computing power and all this knowledge
about neuroscience, let’s model the brain. And the premier example of
this is something called the Human Brain Project
which is centered in Europe and throughout Europe, and
it’s very, very ambitious. The Human Brain Project is
trying to model a brain, all the way down from ion
channels and synapses and neurons and the entire brain and the
entire structure of everything. That’s if I’m left here
for different scales, this is one of their brochures. All of this in scales in which
they’re trying to deal with. That rosette in the center,
which you can’t read any of the details on, those
are, each one of those leaf nodes on that picture is
a scientist who is one of the principal
investigators on this project. You can see there’s hundreds
of them, hundreds of scientists in all these different
fields are coming together to build this monster brain simulator. Now on the right is a picture
of one of the early models they did from when they
were working with IBM called a Blue Brain. These are cool. I went and saw this. There’s these walls of pictures,
you’re surrounded by walls of these neurons are
being projected and spikes are going around. It’s very, very impressive. Now, it’s interesting
when you think about this, and they won’t disagree with this. This is not a criticism,
they want this description. There’s actually, first of
all there’s no theory here. There’s no expert, they have
no idea what this should be doing, they’re just
making a software simulation of millions of neurons. And they’re turning it on
and saying what happens? Well, they might be learning
something from that. But there’s really no theory. And I just gotta believe you’re
not gonna make any success if you have no idea what
this thing should do, or what pieces are important,
what pieces aren’t important, what things are essential,
what things are not essential. You’re just not gonna get it right. It’s just impossible, and
that’s my opinion about that. And they don’t really disagree with that. The other thing is, there’s
really no attempt at machine intelligence, we realize it’s
not gonna happen this way because they really don’t
know what you should do and how it’s gonna work. But it’s interesting. They’re not viewing this
as a way of saying okay, we’re gonna model the brain
in this huge simulator, we’re gonna learn all kinds of
things maybe about diseases, it’s gonna be a tool for
scientists, and maybe we’ll figure out how intelligence comes
out of this a little bit down the road. So, that’s pretty cool, very interesting. A lot of money and effort
being spent on these kind of things. This is not the only initiative,
there’s some other ones that are similar to this, but
this is the one that is most in the news these days. Now, you might not, now I’m
not a big fan of all of these as a way of getting to
machine intelligence. I think we need a different
way of thinking about it, a different way of looking at it. And, you shouldn’t be surprised. You know, what would by my
ultimate approach to this? Well, you’ve already seen it,
it’s this stuff right here. My alternate approach is you first start with the neuroscience,
start with the brain, use that as a set of constraints,
build those principles, and then once you have those principles, then you can start implementing. My idea here is if I want to
build intelligent machines, it shouldn’t be performance based. Think a mouse is an intelligent
machine, intelligent animal. A cat, a dog, a monkey, humans,
these all have different levels of intelligence. It’s not that surpassing the
intelligence which we want to do, but if the goal is we
should be working on the right set of principles. If we realize what those
principles are, then we can say this machine is intelligent. It may not be very
intelligent, or it may be super intelligent, but it’s the
principles that count. And, that’s the approach I want to do. So, I want to go through
that list of principles now. For those that were here
yesterday, I talked about a few of them, I’m gonna add some more today. And I’m gonna review a few
of the ones that I talked about yesterday, so. And I’m gonna end up a little
bit of review for a few minutes, and for those of
you that saw this yesterday, think of it as a refresher course. But it’ll be much quicker
than what I did yesterday. Okay, so one more thing
is you have to understand what the neocortex is doing. The neocortex is a memory
system, it’s not a computer. It’s a memory system. It stores information in it’s
connections, and it builds a model of the world. When you’re born, it
doesn’t know about anything. It doesn’t know about all
the objects in the world and how they relate to one other. It doesn’t know about
presidential debates and computers and rooms and so on. And it has to build the knowledge of that. And it does that through its senses. So we have arrays of senses
here, the retina and the cochlea and the somatic sensors
of the body senses. Those are millions of sensory
bits coming into the brain, streaming in real time, high
velocity, and very rapidly changing; the brain gets
these in and it has to build them in from the world,
that sort of thing. In the model, it has several things. It makes predictions about the
future, it detects anomalies, and it generates actions. So, the exa– Give you an example, I
talked about this yesterday, but you know, your brain is
constantly making predictions about what it’s gonna see, hear, and feel. You’re not aware of most
of these predictions. They occur, it’s all
subconsciously for the most part, but you know when your
predictions are violated that something’s different
than what’s expected. The example I used in my book,
I was sitting in my office one day and I realized
there’s all these objects in my office! And yet, if any of them,
if some of them just moved a little bit or disappeared,
or some new object would appear I would notice it right away. But I’m not sitting there going, let’s take inventory of my office. I’m just looking around and things happen. Similarly, your expectations
when I speak, you have expectations what you’re gonna
feel when you touch things, you have expectations
what you’re gonna see. This is proven, we know this is a fact. Your brain is constantly
making predictions. So, this was a way of figuring
out, a cornerstone for me, like, how is it, what kind
of memory system would make predictions, and the answer
to that is a memory system that learns temporal kinda,
what normally follows what in various ways. Okay, so that’s the basic idea. Let me just go through, I’m
gonna list these principles the neuro cortical function
that I think are essential for true Machine Intelligence,
and then I’m going to go into detail about a couple of them. So here we have this cortex,
we have this high velocity data stream coming from a set of sensors. One of the action bits that I
think are really gonna define Machine Intelligence. The first one you might be surprised. You certainly would have
thought about this much, but when you think about
intelligent biological systems, they all have these sensory arrays. We don’t have a camera for eyes,
we have an array of sensors and they are processed like a
whole, like millions, millions of sensors. There’s nobody looking at that
entire picture in your brain. And so, the same with
the cochlea, and the same with your skin. And so, we now understand
this is an essential property of how the memory system
in your brain works, and I don’t think you can get around it. So, Intelligent Machines are
going to have to have sensory, low level sensory arrays. You just can’t go, no one
just pumps Shakespeare into your head. You know, that doesn’t happen. You read it in a very complex
thing seen through your eyes. Or you will hear it being spoken. Or you can feel it through Braille. But the point is, it’s
gotta go through this array of sensors through time
to get it in there. The second thing is that the
neocortex is a hierarchy, and I talked about this
a little bit yesterday. When we look at it, physically
the memory in the neocortex’s sheath is arranged in
a hierarchical fashion. These regions are connected
to each other in a hierarchy, and the information flows
up and down the hierarchy. This is a physical fact about
the brain, and this is a very interesting observation about
the kind of memory system it is, it is a hierarchical memory system. And, you can’t get around that. Intelligent Machines are going to have hierarchical structure. The next thing is, and
this is part of the theory that we’ve developed is that
each reach in the hierarchy is doing is a form of sequence
memory, or different type of sequence memory. And so, as you are, when
you’re hearing my speech, and you’re hearing these
complex patterns coming in through time, the way you
recognize my speech, ’cause you have memory of what words
sound like in order, and what sentences and phrases sound
like in order, and what things typically follow what. And so to recognize speech,
to recognize when you touch things, and even vision is a
hierarchical temporal process, and so there’s this sequence memory. And the idea of sequence
memory in a hierarchy is that you’ve sequence memory at
the first levels, recognizing very small patterns in space and time, and then they collapse
into longer patterns of space and time and so
on going up the hierarchy. And you get to the top of the hierarchy, you have representations of higher level objects in the world and how they behave. Some of this is very well understood. Some of it’s not understood at all. But, this is a fact. And, one of the things
about the hierarchy, which I mentioned yesterday
and haven’t mentioned yet today is that all the regions
of the hierarchy are doing the same thing so once we
understand you know, how one part’s doing it we understand
how all part’s doing it. So, these are attributes
that an intelligent system are going to have to have. The next one is Sparse
Distributed Representation. This is the language of the
brain, and I’m gonna go over a bit more of this in a moment,
but it’s the way when we, everywhere we look we find a
few things that are active, few cells that are active, and
most cells that are inactive. And they’re properties of
this which are important that you’re gonna need to understand. And, I contend that no
intelligent machine, biological or otherwise, can work
without Sparse Distributed Representations, so we’re
gonna go a little bit more into that. And then, here’s some
things I did not talk about much yesterday. Everywhere you go in the
neocortex, there’s these cells and sequence memory that
are learning the patterns of the world. But everywhere you go,
no matter where you look in the neocortex, you see
outputs that are descending to motor parts of your brain. And, everywhere is that
sensing information is also directing behavior. You can’t separate out behavior
from sensory information. You can’t separate out
inference from behavior. And think about it, when
you, when you interact with the world, whether
you’re just moving your eyes or walking around or touching
things with your hand, you’re changing what your
senses are gonna feel. You know, I sense my fingers
aren’t just gonna feel this podium because they happen
to be feeling the podium, it’s ’cause I have moved my hand there. And there’s a tight coupling
between your behavior and what you sense. This is the sensing motor
integration issue, and we have, this is part of how our brains work, and this is an essential
feature that our Intelligent Machines have to have. Then there’s an attentional mechanism. I’m giving you the laundry list
here, okay, so bare with me. I think this might be the last
one, or maybe have one more. Attention mechanism, in the
hierarchy, there’s information flowing up and flowing down,
but we know that there’s actually ways of turning it off. And so, you can attend to some
subset of your sensory stream so you might be sensing,
you might be zoning out on your not really feeling
too much right now, you’re listening carefully to my words. Or maybe reading, you’re
not listening to my words, and you’re paying attention
to the stuff on the screen. But I could also tell you
like, okay look at the word principles at the top of the slide. Now, there’s a pattern coming
in your brain right now when you’re seeing the word principles. And I say look at the I. The letter I. The same pattern’s coming in
on your eye, but now you’re focusing your attention,
you’re attending to a subset of that information. Same information coming in the
brain, and you’re attending to a subset. Now you say look at the
dot on top of the I. Focus on that. Same pattern coming in
your eye, the whole picture but you’re now attending to
a smaller piece of the image. You’re doing this all the time. You’re not aware of it. You’re doing this all the
time, tuning in some parts and taking out certain parts
and this is pretty important to have a complex sensory environment. So, this is going on in the brain. I’ve showed you in this picture
here where I say hey look, I can turn off those little
X’s, we can turn off some of the input and focus on
other parts of the input. And this is a pathway through
the thalamus which does this. This is very important part
of how we interact with a very complex, rich world, and is an important part of intelligence. Now a few things on this,
there’s a lot of things that aren’t on this list I
just mentioned a couple things right here. One is, you notice I
don’t mention emotions. There’s books about emotional
intelligence and so on. I don’t believe you actually
have to have emotions. You have to have emotions to
be human-like, you have to have emotions to pass the Turing Test, but to be purely intelligent
to build a model of the world that’s very, very sophisticated
and build predictions, and detect anomalies, and
discover structure of the world, you do not have to have emotions. Emotions are all about, it’s
all about an ubering system. It says you know what,
this is really dangerous, or this is really good. I need to remember that. Or I need to avoid that. But you can do an awful lot without that. And I do not believe this is
something that’s essential for intelligence. And finally, you don’t need
to have a human-like body. This isn’t about building
robots, and you know, and yeah, it’s not about building robots. And, you know, you want to
have a human body-like body, but it’s not about that at all. We can have intelligence
embedded in all kinds of systems, you might not even be able to see it. It’s a bunch of computers
running some place with some sensors on their own. So we don’t need those
things, we’re gonna get rid of those things. We’ll just stick to these principles. Okay, so, by now there’s two
here that I told you about yesterday that are
really, really important. And so, what I’m gonna do is
I’m gonna go through the slides I had two slides I had yesterday
about Sparse Distributed Representations, I’m
gonna do those completely. Because to me that is the most
important thing, if you want to operate anything, you have
to understand what Sparse Distributed Representations are. And the second one, I’m just
gonna give you a cursive review of the sequence memory. And I apologize for those
of you who totally got it yesterday, and say why
are you repeating that? But I doubt anybody
totally got it yesterday. It takes awhile. Okay, so here again, the two
slides on Sparse Distributed Representations– It’s easiest to do this when
you compare it to a computer. If you know how computers
work, you can easily see how this works in the brain. And I’ll do this a little
bit quicker than usual. In the computer, we have what
we call dense representations. We have 8-bit, 16-bit,
64-bit entities, you know, 64-bits at a time or something like that. And we use all combinations
of ones and zeros. So if I have an 8-bit quantity,
a bite, all 256 possible combinations of ones and zeros
are used to represent things. That’s why it’s called dense. There’s an example of the
asking code for the letter M. It’s just some arbitrary
assignment of ones and zeros that says this is M. If I ask you what those different
bits mean, in the letter M they mean nothing. In fact, if I change one of
the bits, I get a completely different letter. So, I have to look at all
of them to get any idea what this is. And actually, nothing about
those bits would tell you what’s M. It’s placed elsewhere in some
table that someone says okay, that’s the letter M. And then, finally, these
representations are assigned. They don’t have any inherent,
you know, they’re not learned. Someone just said here’s the
ASCII code, we’re gonna use this for the next 100 years. In the brain it’s very different. In the brain we have cells,
and if you look at the cells very few are active at any point in time. Most of them are inactive,
or relatively inactive. And so it’s like, and
there’s thousands of them. And so in a Sparse
Distributed Representation, we can represent the
cells with zeros and ones, and we can say that those
are several thousand, or tens of thousands of these bits. We don’t, we typically use
things about 2000 bits long in our work at Numenta,
and I’ll talk about that in a little bit. And, they’re mostly zeros. But we also typically have 2 percent on. So have 2000, I have 41-bits,
and I have 1906 zero bits. The, there’s nothing
magic about those numbers. But sparsity is important. Most bits have meaning of some sort. I can, you know, they
don’t change over time. I say here’s a bit, it was
like here’s a cell to cell represents something that
doesn’t arbitrarily change from moment to moment. So this cell represents
a line in a certain part of the visual space, it’s
always going to represent that line in that part
of your visual space. If this cell represents a
part of recognizing a face of a human, it’s always
going to be that way. And what the basic idea here is, when you form a representation,
if I have 2000 bits, I’ve 2000 somatic meanings,
and I pick the top 40, the top two percent that
best represent this thing. The example I used yesterday
is I wanted to represent a letter, I might say here’s
bits, and we wouldn’t do this, this is an example, I
might have bits for okay, is the letter a consonant or a vowel, is it sound o sound or e
sound, is it hard or soft, where is it in the alphabet,
does it have ascender and descenders, and so on,
and I pick the bits that best represent that thing. And now my representation
actually tells me what it is. There is no external place
where I have to say oh, this code represents X. It’s right there in the code. If I know what those bits
mean, the encoding tells you what it is. That is the entire definition. And so, if each bit has
semantic meaning, and these bit definitions have to be learned. Now there’s these properties
that are SDR’s which are very, very important. One is one of similarity. If I take two Sparse
Distributor Representations and I compare them bit by bit,
if the shared bit, that means they share semantic meaning
of some sort, and therefore they’re similar semantically. The more bits they share,
the more semantically similar they are, very simple property. The next one is if I
want to store a pattern, a sparse representation, and
I want to recognize it again. So, here’s a pattern, I’ve
seen this now, I want to see this accrue again. Well, I can save all 2000 bits,
that’s one way of doing it, that’s how a computer programmer
would do it, but we can do something simpler. We can just say well this sort
the locations of the ones. So I have 40 one bits, I’ll just say where are those 40 one bits. And if I see ones in those
40 locations, I know I’ve got my pattern because that’s
all there is is 40 ones. And then we can ask the next
question, which is what if I couldn’t save all 40 indices. What if I can only save 10 of them? And so I say you can’t store,
we’re just gonna have to randomly sample the 40 and
just save the locations of 10. Well, so I do that. And I say sure I’m storing
some of them, but I’m not storing other ones. Now a pattern comes in, and
you say is it the same pattern or not, and I say well
look, the 10 are the same. What are the changes of the
other 30 being the same? You say well, I could be wrong. Those other 30 could be some place else. The chances are very, very, very unlikely that’s going to occur, but
if it does occur, you made a mistake, but it’s a mistake
for something semantically similar to the thing you did store. So it’s like okay well, I made
a mistake but it’s similar to the thing I stored before, and that’s often very good enough. And then finally, this is
the most difficult property, is one of union. And you can take these Sparse
Distributed Representations and what if I took 10 of
them, and I say order them together, so I have these 10
patterns, each of two percent of the bits on them, I
already have one pattern, which is about 20 percent of
it was, of the bits are on, and I have this union. And I say well, can I tell
you what the first ten were? No, I can’t do it. I cannot undo this operation. But I can do something very interesting. Is I can take a new Sparse
Distributed Representation, and ask is it one of the members. And the answer is, I can do that. I find that the patterns,
the ones in my unknown match the ones in the union, I can
be very certain that this thing is a member of the original 10. And you might again say well,
it could make a mistake. But simple math will tell
you it’s not gonna happen. Very, very unlikely. And this is, I use this when
the brain makes a prediction, it’s making predictions
in the activity of cells, and it essentially says I
can have multiple predictions going on at once. I can predict many things
that might happen next. And I can tell you if what
actually happens was one of those things. So when we make predictions,
you often can’t tell me exactly what you’re predicting. You don’t know. But when something
unexpected occurs, you know. And that comes from that
property I just showed you there. ‘Cause again, this is
the most important thing about Machine Intelligence
and about brains. You want to walk away and
say hey I heard this guy Hawkins talk, he’s kinda crazy
but one thing I remember, this would be it, brains
and brain machines are gonna be based on Sparse
Distributed Representations. You can bank on that. And many people will in the future. Okay, now I’m gonna talk very,
very briefly about sequence memory, I will not be able
to go into all details about this, but I want to
give you the flavor for it. Yesterday I gave you
many more of the details. Here we’re seeing a Sparse
Disubstituted Representation, but instead of showing you
ones and zeros, we’re showing it as little cubes. We can imagine those cubes
being cells in the brain. And, the red ones are active and the white ones are inactive. And at any point in time,
I’ll have two percent of them active like this, and then
at another point in time we know they’ll set like that. And the basic idea we want
to do when we want to run sequences, is we don’t
try to learn the sequence of everything, but every
cell in itself, every little cube here tries to
predict its own activity. Tries to learn when it
follows something else. And, if you do that, you end up with a
Distributed Sequence Memory. So, when a cell becomes,
so we have these patterns coming into your brain
right now as I’m talking, these patterns are flashing
back and forth like this all the time, and when a
cell becomes active it looks for cells nearby, it doesn’t
have to look into all of them, it just has to subset a few
and say memory I remember which ones are active. Those are like those indices,
which ones are active? And I’m gonna save those
ones, and know if I see them happen again, that I can
predict my own activity. And, here’s a situation where
I have an input coming in and a whole bunch of cells,
yellow cells, predicting they’re going to come next,
just to be in a situation where I’m predicting multiple things. Like if I had A followed
by B, and A followed by C, and A followed by D, I
show you A is gonna predict B, C, and D. Then I would show yesterday
that this is a first order memory, meaning it can only
predict based on the last thing that happened. It has no history. It can’t tell you things
about how long in the past. Imagine a melody. A melody is a high-ordered pattern. Think about Beethoven’s Fifth. It goes bum-bum-bum-bum,
bum-bum-bum-bum, bum-bum-bum-bum bum-bum-bum-bum– The first four notes,
bum-bum-bum-bum, are repeated as the ninth to twelfth notes. Exactly the same notes,
but you don’t get confused. And, I don’t get confused as
it goes through the entire melody, I never get lost and
say oh, it’s the beginning again, starting over
again, starting over again. In order to do that, you have
to have a high-order memory. You don’t want to get confused
if the beginning only sounds like the beginning, and you’re lost. So, we need to do this and
without going into the details, the solution to this problem,
I believe, has to do with using columns of cells we
see in this in the brain. Instead of one cell for a
bit, we actually use multiples bits per, we use multiple
cells per bit in our Sparse Distributed Representation,
and this gives us this very, very high capacity memory. And I want to walk through it here. I’m just gonna tell you
that it’s the variable order sequence memory, it
does multiple simultaneous predictions, it’s extremely high capacity, it’s a distributed memory
system meaning it’s fault tolerant, you can drop out
cells, and neurons, and columns, it’s just like in a real
brain, it keeps working, and it does semantic generalization. If you really want, if you
missed yesterday’s talk, or if you had yesterday’s
talk, you can read the full details of this, there’s a
live paper on our website which tells you all about
this in great details. So you can get into well
what was he talking about, you can read about it. Okay, so that’s it for my review. And now I’m gonna go forward. We’re back to this situation right here. You say these are my six
attributes, I believe need to be a part of my intelligent machine. Now, you know, intelligence is a scale. It doesn’t, you know, there
isn’t some threshold to it. As I said, we have lots
of intelligent animals with different capabilities. And so, maybe you won’t be
able to do all of these things. But a really intelligent
machine would do all these. Where are we today? Here’s where we are today. Today we understand the sensory
arrays and streaming data. We are modeling this today. We understand the
sequence memory very well. We understand the Sparse
Distributed Representation very well. We partially understand the hierarchy. We have some pieces of
it, but not all of it. So some of it we’ve done some
simulations with hierarchy, some simulations without hierarchy. There’s more work to be done there. Now I’m going to show you what
can you do with a very simple version of these, these capabilities. Can I do something, I’m not
gonna call it an intelligence machine, but it’s on its way to being an intelligence machine. We’re taking baby steps
there, but turns out I can do something very, very
commercially useful with just these properties. And that introduces the
concept of our product. And to tell you about our
product, I have to give you a little bit of a side version into data. I’m gonna have to talk about
world data a little bit. So bare with me, it’s not
gonna be very technical at all. Today, we have the ability to
collect huge amounts of data. Perhaps you’ve heard about
Big Data, it’s a popular term these days, and we’re storing
so much data that it’s sort of like the Library of Congress
is the metric for how much data we’re storing. Oh, we’re storing four
Libraries of Congress an hour or something like that, you know. Huge amounts of storage. And the problem is what
do we do with this data? I mean, we put it in databases
and people can look at it. Tools of visualizing, and
so on you see the patterns, and the build predictive models. They hire machine rental
is suppose to come in and they build these models,
it takes them months, and then the models get out
of date and they do it again, and so on. This is an un-scalable problem. It’s not, this is not the
solution of the future. There’s all these problems with it. There’s problems with preparing the data. There’s problems with the
models becomes obsolete. The patterns in the world
change, and someone builds a model and a lot of
them aren’t good anymore. You see this in credit card fraud. People are always, there’s
the fraudsters, that’s what they call them, and then
there’s the people who build the fraudsters detector. And so they build these models
to detect the fraudsters, and they say ah, the
model’s working really well, but the fraudster goes out
and in a month they’ve got new ways of cheating. And it’s just a race that
goes on and on all the time. And there’s another problem
with this is that it tires people, lots and lots of people,
machine experts, and so on. So, this is not a scalable solution. And by the way, this is
an important problem. The world is going to be awash in data. We are going to have
trillions of data centers. You may have heard of
the Internet of Things. Everything in the world is gonna
be connected, and streaming data someplace. What are we going to do with this? Not this! Okay, the answer to this is the following. The answer is, you’re gonna
take that data and you’re gonna stream it to online, what
we call online mouse. There’s a continuously learning
models, and you’re gonna take actions directed from it. You’re gonna take the data,
put it through these models, make predictions, detect anomalies, take actions, immediately. Does this sound familiar? This is what brains do. Streaming data, continuously
learning models, making predictions, detecting
anomalies, and taking action. And so there’s an opportunity
here to rethink the way data is acted upon in the world. And this turns out to be a
very powerful idea, and we’ve built a product called
Grok which does this. The key criteria here is
the automated model creation and the continuous learning
because as the patterns change in the world, the
models have to adapt. Just like, you know, as
a human you do this too. Everyday you learn something new. You’re adapting, bend some
things you’ve learned in the past and you’re adapting new patterns. And so, you do this continuously. And it’s very, very important
to find the temporal and spacial patterns in the data. Very few people look at the
temporal patterns in the data. That’s what brains do. But the temporal patterns
in the data are very, very helpful. They tell you what’s gonna happen next. They tell you when things are unexpected So, this is a great
model for what brains do. So we’re gonna build miniature
brains, we are doing this! Build limited brains, if you
will, that take data streams and make predictions and take actions. So, building a product
is quite an endeavor. And so, I’ll just walk you
through some of what this is. Here’s a diagram of how we do this. On the left, I have these
three vertical bars. They’re representing records
of data from some data stream. You can imagine this coming off
the same computer or server. You can imagine it coming
off sensors on a building. And these are records coming in in time. They may be coming in very
rapidly, once a minute, once every five seconds. They may be coming in
slower, like once an hour. They have multiple fields. They might be numbers and
categories, and things like that. But we feed them into our system. The first thing we have to
do is we have to turn these inputs into Sparse
Distributed Representations. That was one of my criterias. The sensors have to produce extremely Sparse Distributed Representations. So we have a way of doing that. I don’t mind telling you about it. But I’m not gonna go
through it in this talk. We can take numbers and
turn them into Sparse Distributed Representations
that have the right properties. We can take categories of
information, like you know, male female, and days of the
week, and things like that, and we can turn them into Sparse
Distributed Representations and we do that. Then we feed them into
this sequence memory. It’s corticolmar, 2000
columns, 60000 cells, you know, 300 million synapses doing
the thing I just talked about before, those columns and all that stuff. And from that we make
predictions and take actions. Here’s what a user would
do with this system. They don’t need to know any of this stuff. The users essentially
says oh, here’s my data, I have a stream of data
coming from this building. For example, I want to find
a problem and try to make predictions every so often,
and here’s what I’m trying to predict, and then our
product Grok basically creates these models, figures out
how to do them and runs continuously, it finds
spatial temporal patterns in the data, and it makes predictions. And it can tell you the
probability of these predictions. And there’s lots of areas
for applications here. Energy pricing, energy
demand, product forecasting, ad network returns, et. cetera. We have all kinds of people
trying to do machine efficiency, kind of predicting which machines to use and when to use them,
and things like that. I’m gonna give you just a
couple of examples so you get a flavor for what this is like. This is not an important detail,
but for the computer people in the room you might care about this. Today, this is implemented
on a cloud server, an Amazon cloud server. Doesn’t have to be. This could be embedded in
things, could be embedded in cars, chips, whatever. But basically today, we’ve
implemented our first little brain models as a service on the cloud. So I’ll walk you through
a couple of simple slides, an easy to understand version
of this which is energy. You may not be aware of
this, but large consumers of electricity, they pay and
they decide how much energy they’re going to use sometimes
on an hour by hour basis. Now there’s a market for like
a large factory when they talk to utilities, utilities
say well I’ll sell you this electricity at four
o’clock in the afternoon at this price if you take so
much, or don’t use so much. And then the consumer of that
energy says okay, I’ll either do that now or not. Then some people they will
pre-cool buildings because they know that the electricity’s
gonna be more expensive later in the day. So there’s this market going
on for larger consumers of power that you’re
probably not aware of. It’s called the Man Response
Market, and if you can make that more efficient, you can
save energy and save money. So here’s a typical example,
here’s an energy profile of a building, it’s some
factory, and you can see there’s a pattern here. It turns out these peaks were
caused by day of the week, so you can see there’s five
days a week and then a weekend comes along the factory shut down, and nothing’s going on there. Now we can feed this kind
of energy thing into Grok, and it can learn this, this
looks like a fairly simple pattern, but it’s not as
simple as you think it is. So, in this case, the customer
and what they want to do, is they said at midnight,
I want you to predict the amount of energy that’s
going to be used every hour for the next 24 hours. That’s their problem,
so we have to do that. And we can do that using
these kind of models. So, that little red line says
look, can we predict that. And so the next slide
here I’m gonna show you, this is just showing in
the red is all predicted in the blue, and the blue is our actuals. And, this is your streams
of data into Grok. Now Grok doesn’t know anything
at all about what this information represents. It doesn’t know if it’s
energy or you know, grams of alcohol. It doesn’t really care. It’s just a number and it
looks and finds these temporal patterns and says oh, I can
see these temporal patterns. And it looks like it’s
doing a pretty good job. It actually is in this case,
although you can’t tell too well just by looking at these graphs. But trust me, the customer
is very happy with this, and it is doing something
very significant. And, here’s a situation where Grok started to make a mistake. There was three days in the
week, and you can see right down here it starts
predicting the next day. Well, it turns out this
was a European holiday, and it hadn’t seen that
before, and it says oh well, oops, oh no it’s not, and it
quickly says no that’s not the pattern. Here’s another pattern. This looks like the one I’m
suppose is, it’s suppose to be like this, and
it quickly be covered. Um, there’s another example. This one looks a little bit harder. So I’m just gonna save you
flavor by looking at it. There’s a company we work
with that’s trying to demand, to predict how much
demand for their service. And they have a service which
is encoding videos on the web. And so customer sends a
video, it has to be encoded in many different formats,
so you can look at it on your phone and other things, and they want that immediate response. As soon as they start
sending that video, they want the encoding to appear on the web. And so this company has
to leave lots of computers around running all the time
doing nothing, because they want to be able to make sure they
can catch a spike in demand. And this graph showed you
that the demand is actually very spiked, goes up and down
all over the place, very hard to see what the patterns in there are. And yet, we can run something
like this through Grok, and Grok will say maybe
there’s there’s patterns in the afternoon, maybe when
school gets out, who knows what, it doesn’t really matter. If there’s patterns here,
I’m gonna try to find them. It can’t do a perfect job. But in this case, it did a
good enough job that they can save about 15 percent of
their cost, which is very significant to them. So, this is the kind of
things we’re applying it to. Now here I’m gonna show you
something a little bit more technical, and I hope you can follow this. On the right, just pay attention
to the thing on the right. I told you that models
have these like 2000 bits, these 2000 columns of cells if you will. There’s 2000 little circles
on that drawing there. And these are representing
the activation, the internal activations of our cortical model. These are like the 2000
columns in our cortical model that we’re running. We’re just looking down on
top of them if you will. And the green dot means,
if you counted them, there should be 40 green
dots there don’t bother, if you count them there’d
be 40 green dots there. These are both predicted
and what actually happened. So, Grok was saying okay,
in the next representation I’m expecting these 40
attributes to be active, and it turns out these 40
attributes were active. It was a perfect prediction. It happens a lot, but not always. Here’s another one where
those little blue circles are things that were
predicted that didn’t happen. But everything that did
happen was predicted. So this is again a
multi-prediction going on. This says okay, I can see
three or four different things occurring now, one of
them actually did occur. That’s good, we like that. So that’s what’s representing there. The blue circles were
predicted but didn’t happen. And finally, here’s a situation
you probably have trouble seeing this, there’s a bunch
of little red circles on here. If you can’t see that, trust
me, there’s a bunch of little red circles, so we have
red circles, green circles, and little blue circles. And what’s going on here
is Grok made a prediction. Some of those attributes
turned out to be true, those are the green dots. Some of the things were
predicted didn’t happen, not a problem. But some things occurred
which weren’t predicted. Some attributes occurred
in the input stream which weren’t predicted, and
those are the red circles. And, what the point of the
slide is to say when you make an error in prediction,
it’s not a binary thing. It’s not one or nothing. It’s a very nuanced thing. There’s somethings that
are right, and some things that are wrong. And if you actually could go
and look and probe into this, you’d find that you could
tell what semantic was correct and what semantically was incorrect. It’s a nuanced thing about
what an anomaly is or what a prediction is and so on. So this is the kind of
stuff that we do internally, and you can use this in various ways. Here’s an example, and this is
my last example in this thing of a windmill. This is one of those
huge windmills off shore, have you seen these
offshore windmill farms? It’s amazing. These things are like monstrous. And the North Sea they have quite a few. And out there in the sea, these
monstrous windmills running 24 hours a day, and they’re
very, very expensive. And if they fail, it’s very
expensive to replace the parts. Like, you know, the gear box
costs several hundred fifty thousands dollars, and probably
cost $100,000 to replace it. So, if they can, if they can
detect anomalies and detect before failures occur,
it’s worth a lot of energy and money and so on. So this blue line happens
to be the energy consum– Excuse me, it’s the temperature
of the oil in a gear box in a windmill in the North Sea, okay? I forget what year, what time
this is, maybe it says it. Yeah, in 2011. This is the last one. So, you can see that temperature
is going up and down, overlapping all the time as
the wind goes and changes. It’s a very complex pattern. And, on the bottom on the
right there, you see sort of an aggregated anomaly score for Grok. Grok is saying oh I’m trying
to predict what’s going on here, I can’t predict all
this stuff, but I’m looking for patterns that I haven’t seen before. And you can see on the down
here, there’s a peak there, there’s actually two peaks
on the anomaly score. And what’s happening here
is, there’s nothing, if you looked at the temperature of
the gear box, there’s nothing wrong with the temperature, it’s in range. But the pattern is wrong. It’s like I’m listening
for a melody, and the notes are in the wrong order. And so, we can say you know what? There’s nothing out of range
here, but it’s not like I’ve seen it before. And it says, you know what, I
think you outta look at this. And it turns out, that indeed,
very shortly thereafter there was a maintenance event. So this is an encouraging type of time. And this is worth a lot of
money and saves a lot of energy and so on in the world. So, that’s the end of
my discussion on Grok. So now I’m gonna talk to the
end of my talk, which is about the future of machine intelligence and why we should do this. Where is this all gonna go? Can we build you know,
really crazy great machine. Well, it turns out, yes we can. I’m absolutely certain we can
build amazingly intelligent machines, but the question,
what are they going to be like? And what are they going to do? Is this good or is this bad? And how are they going to be amazing? What’s unusual about them? And so on. So, let’s just walk through
some of these things. You know, there’s two basic views on this. One view is this is bad. You know, Skynet, I’m not
a Science Fiction fan, but I know what these are now,
Skynet is this bad machine intelligence that takes over the word. So there’s the Matrix. It’s like we’re all being plugged
in, we don’t even know it. We’re being consumed for
food or something like that. There’s the Terminator I,
that was the bad Terminator. You know, robot guy. Then there’s the benign view. You know we all wanna have
CP3 around our house, helping us out, do things, or maybe
we’ll all being playing games with gloctomite things in the
future, or maybe we’ll get our entertainment by donning one
of these hats and sitting back and going wow,
that’s great, who knows. Then there’s, the ambiguous
thing in the middle. Well, we thought it was good
and it turned bad, you know. That’s how. But let’s just talk about my
opinion about these things. I’m gonna talk about some
things that are definitely going to happen in my opinion. Absolutely gonna happen. Number one, we can make
artificial brains that are faster than biological brains. Biological brains are
really slow, actually. Neurons can’t do anything
faster than five milliseconds. That’s like the flaw. But we can do things in
silicon that are million times faster than that. So we build machines with
those principles, we can make brains that are millions of
times faster than humans. I’m sure this is gonna happen. And I have all kinds of
interesting ideas about how would you take advantage of that. Certainly you could get to
conclusions quicker, but more importantly can I deal with
things that are very high velocity data streamers. Can I do things that are working non-stop. Can I have a thousand physicists
working 24 hours a day on some problem that would
take a thousand years to figure it out, type of thing. I think that’s gonna happen. The second thing, there’s
no reason why we can’t make artificial brains
Intelligence Machines that are bigger, not physically bigger, but bigger in their memory capacity. Bigger hierarchies, more
regions, more size, and so on. Why not? Our brains are constrained
by the birth canal. You know, we’re pushing the limit. And we have a, naturally we
have a high death rate in birth of humans because of this problem. Nature might want to build
us bigger brains, but they don’t come out. And so, you know, we
don’t have that problem. We can make them as big as we want. And they will have more knowledge, and have deeper insights. We don’t know yet, but I’m
certain this is going to happen. Another area that’s very, very
interesting, we don’t have to stick with any kind of
senses that humans or other kinds of animals have. We know first of all there’s
a diversity of sensors in the biological world,
we just have one set, but you can imagine senses that are huge! Arrays that cover entire planet. You can imagine senses
that measure things really microscore, microscopically,
nano-sensors that are looking at protein folding and things like that. And then, these artificial
brains, these machine intelligence will be
able to understand worlds that we have trouble understanding. Everything a human has to
understand, we have to put into something that runs at our
speed and through our senses. We have to come up with
visualization tools, so if I’m trying to
understand protein folding, I gotta have visual graphics and so on. It’s not a very good fit. But if I had sensors that
really could live in that world, they would think in that world. We could have 3D sensors. Right now our sensors are
all 3-dimensional seeds. The theory says that there’s
no reason you can’t have higher dimensional sensor arrays. So that’s really cool. I think a lot of that
stuff’s gonna happen. You know, fluid robotics. Today we have no fluid robotics. If you look at the robots today,
they’re so clunky and slow and difficult. We’re not even close to this. But we will have fluid robotics
based on these principles where machines can operate. Again, I’m not saying they
look like a human necessarily, but they’ll be able to
do things very carefully. And I’m sure we can create
intelligent machines that will explode the universe for us. Okay, and another interesting
idea is the whole idea of the distributed hierarchy. Now I told you, you know,
we have this hierarchy in the brain, in our
heads and inner cortex. And that’s how everything works. But what if I could distribute it. What if I could have parts
of the hierarchy all around the world? And then they combine together and so on. I can model very, very large systems. These are ideas that who
knows which of these are going to take place, we don’t really
know where these are going to go, the history of
technology is it goes in ways we just don’t anticipate. And often, to tremendous
light and benefit for us. And who could have anticipated
where computers were going to go, where they are today,
with GPS and cell phones in your pocket? 50 years ago, no one
could have imagined that. Who could have predicted
our communications abilities starting with the telegraph,
what we can do today? Unbelievable. And so these are going to go
in these directions as well. Here’s some things that might
happen, I don’t really know. I’m kind of on the fence about it. One is, humanoid robots. Are we really gonna create CP3O’s? I’m not so certain. You know, I suppose if you
really, really wanted to, I suppose you could, but I’m
not sure we really, really want to. This is not about you
know, science fiction. This is about building tools
for us as humans to make our lives better and discover things. And so will this happen? I don’t know. If you want to do that, you
will have to build a lot of things that I didn’t talk about today. Probably have to have emotions,
you’d have to have bodies… Oops. You have to have bodies and so on. So maybe that’ll happen, maybe not. Who knows? I’m not so certain about this
computer/brain interface. Now this is really cool work
being done in this area, some here right at Berkeley,
where you know, for people who have damaged their nerve
system, or something like that, we can put patches on their
brains, and they can learn to control things. We already have artificial
cochleas that work very well. And so the idea that we
might create an interface between our brain and the
rest of the world, you know, certainly it’s going to
happen at some level, but I’m asking the question like, are
we going to plug ourselves in in the evening and you
know, be totally wired? I don’t think so, but
I’m not going to discount it completely. You know, who knows. It’s very difficult to tell these things. Here’s some things that I don’t
think are going to happen. I really just don’t. I’m very, very doubtful
this is going to happen. This is a popular thing these days. Ray Criswell talks about this. Like, you’re gonna upload your brain. So, here I am, say okay, Jeff
Hawkins is standing here, I’m gonna take all my brain
connections, I’m gonna stick ’em in some artificial
machine over here, and I will transfer myself from here
to here, and I will have immortality, or I’ll have super powers or something like this. I believe this is a fantasy. And, theoretically it’s possible,
but if you know anything about brains and how they
really work, I don’t think it’s possible. I don’t think we’ll ever get
the technology to do that. And, it’s also, I think it
would be a very unsatisfactory experience. Now imagine I was sitting here
and we were able to transfer my connections into my little
machine doppler over here. And, and I flipped the
switch, lights flash, and bingo it happens. Now, this guy over here says wow! Jeff, I’m Jeff Hawkins, I’m over here. But I’m still here. I says no, I’m still here,
I didn’t go anywhere. You know, we can get rid of
you, the biological Jeff, because we don’t need you anymore. I’m like whoa, don’t do that! You know, I don’t think
this is kinda weird stuff. So I don’t think this is gonna happen. I don’t think we’re
gonna have evil robots. This is another science fiction fantasy. Every time a new technology
comes along people imagine how it’s gonna just destroy the universe. And so, it’s not gonna happen. Someone has to go really,
really out of the way to make this happen. I tell you, the only thing
that could be really, really dangerous if self-replication. That’s the dangerous thing in the world. Machines that can self-replicate. Or anything that can self-replicate. Viruses, and so on, so as
long as you don’t make our intelligent machines
self-replicating, which is a totally separate field and I’m not gonna go there. Then, we don’t have to
worry about hero robots. They’re not gonna be feeling
imprisoned or any of that kind of stuff. Finally, I’m being realistic,
it’s not only gonna be for friendly purposes. So my point of this double
negative here, whatever it is, is that sure the military’s
gonna try to do things with intelligent machines. That happens, they do the
same thing with you know, cell phones, and radios, and so on. So, it’s not always gonna be
for friendly, benign things. But I don’t think it’s a
threat to humanity in any central way. So, I’m not really worried
about this, these bad things. I don’t think they’re gonna
happen, and I think history is on the side of that. So we’ll just get rid
of those guys like that. Okay, now I’m going to finally
talk about, this is my last slide, and we’ll talk
why should we do this? Why, you know, why should you care? I actually think it’s essential. I think it’s essential for
the survival of our species, and I think it’s essential
for our mission as a species if you will. The purpose of life, and so on. So, the number one thing
here is that we can make our world better. Every new technology can
make our world better. And I think having machines
that can help us be more efficient and make the world
safer, and produce better communications, and essentially
accentuate our lives, is essentially the same
way computers have made our lives better. And you know, computers are
a big plus over how things were before. And, there are some downsides,
but mostly it’s a big platform, glad we have them
to do things we need them to do when they’re there. And the same thing will be happening here. We’ll become very reliant
upon artificial machines, not again robots, not
you know, evil things. Just machines that work on
the principles on your cortex that are helping us figure
out how to run the world and so on. But more importantly, if you
think what is the, what is the purpose in life? I don’t know, but maybe
there is no purpose in life. But there is something I know. I know that we are as a
species have been discovering what the universe is. Our path has been to
discover more and more about the universe. And perhaps as we discover more
and more about the universe, we will discover why or how
the origins, we might come to closure about that. We might not. We don’t know. But I don’t know anything
else that’s worth doing. And I think that’s what
motivates many people of science, and the thing is to
everyone at some extent, we want to know. We want to know more about things. And today, our brains are
how we figure out more. And we have scientists who do this. This is what they do. They look at patterns in the world. They say can they discover
the structure and they look at data in the world, they said
can I just see the structure and patterns of that data. Let me test my model, and
build a model of that, and test is done, and I can
assume, and we build models on top of models. Well, that’s what brains do, right? And that’s what science is. And by having very, very
intelligent machines in ways that I can’t even imagine
yet, I am sure that we will be able to discover
more about the universe. I doubt that humans are ever
going to go explore into space and spend years and years
traveling around the universe. Maybe it’ll happen, but I doubt it. But can we send intelligent
machines in our place to go out and discover the
world, come back and tell us about it, yeah, we can do that. So, can we have those thousands
of physicists brains working around the clock? Yeah, we can do that. So, I think this is going to
be a way of really accelerating the accretion or the assimilation
of knowledge and data in the world that’s
gonna be unprecedented. The human brain has been
unprecedented in the long history of biology and our ability to do this. But I think we can
celebrate this dramatic. So I find this all very exciting. I want, you know, this to
me, this is the future. This is so cool. I won’t live to see most of it. But I think it’s worth working on. And I think it’s very
exciting for the future, and that’s why I come and
do this, and come to speak to people like you. So that’s it. Thank you very much. (audience clapping) Let me repeat the question
because not everyone else heard it. So, the gentleman pointed out
that in the model I presented there was no sensory motor
integration going on. There was no action in the model itself. And remember all the
attributes that I talked about, and the sensory motor
integration was one of them, I didn’t put a green dot next to that one. So, that’s an area where we’re
working on, or I’m working on right now which is very,
very interesting and enticing. And, to understand how that same model– So what we know in the brain,
let me tell you what we know about in the brain. Everywhere you look in the
brain, and in the cortex, you look anywhere you’ll
see those layers of cells sort of forming sort of sequence memory. Everywhere you look, there
are cells that project some motor sensor of the brain,
a motor section of the body. So, everywhere you go
there’s a motor output. So we know that this is,
they’re tied together. And, we have a lot of
interesting clues, a tremendous amount of clues, in the
neuroscience literature, about say how do I understand it? What is going on there? What is the theoretical principle
underlying how an action affects the input, and so on. I don’t have the answer to that. But, we’re, I’m hot on
the trail, I’m excited about the progress I made on
it, and maybe in a year or two I’ll come back and tell you about it. In the meantime, the question
that we had at Numenta was can we build something
useful without it? ‘Cause like if I can’t build
something useful from that, Numenta would still be a research company. And we’d be working on it. But it turns out we can. It turns out the problem they
talked about there are simple enough that we can do them
without the motor sensory loop in some sense. And so, I wouldn’t, I
had those six attributes, and I had only three of
them actually in that model. So, there’s a lot of things missing. But the question is can we get started? So, good observation, it’s not there. But I, it looks like we can
go forward in getting started and we have a lot more to do. – [Male] I notice you had,
what was it, 2K cell, or 2K column, 6K Cells, that’s
like 30 cells a column, I know the brain has like
seven in the cortex– – Are you saying does
that match up to what the biology says? – [Male] Yeah. – Alright so, it’s a good
question, although your facts are a little bit wrong there. The neocortex, everyone
gets confused by this. The neocortex has essentially
five layers of cells, okay? Now, I say five layers
of cells, that doesn’t mean five cells. That means five dense layers,
really dense cellular material and the columns span
across all those layers. People say six layers, but
one layer is not cellular, it’s acellular, doesn’t have cells. So we have five of the cells
and the columns go across all of them, and I was allowing
just one of those layers. Now if I look at the entire
column across the whole neocortex, in a real neocortex,
there’s about, depending where you look, but many
places is about 100, 110 cells in a mini-column. We’re not talking about
the ice cube column, but a mini column in the neocortex. And so if I were to say I’m
modeling one layer of cells, I could layer three, which is
one of the prominent layers in the neocortex, and I
say there are 30 cells in a microcolumn, that’s
a realistic number. That’s right in the ballpark
of out of the 110, maybe 30 might be allocated there. The model is not particular about this. I could have 20 cells,
I could have 10 cells, I could have 50 cells. They all works, it’s just brain capacity. So, and then, and so
if I have 2000 columns, this is very, very small
part of the neocortex. 2000 micro columns, these
columns are only 30 microns wide. So, to put 2000 of them,
little tiny of those things in the cortex, one layer,
we’re just modeling the tiniest little part of it. But, those numbers are realistic. I didn’t go through, yesterday
I had a slide where I talked about the other numbers where
there’s about 128 dendric segments per cell, right
in the ballpark of reality. Each cell is something
around 5000 synapses, right in the ballpark of reality. In fact, if you know
anything about neuroscience, a typical cell in the
neocortex is several thousand synapses, there are no
other theories that go what are all those synapses doing. There is a very concrete
theory explains all that. Anyway, so my point is, if
you, if you dive into it a little bit, the numbers
match up really well actually. – [Male] Alright, thanks. – Well that was a good enough
question, wanna ask one more? – [Male] Yeah, please. – Okay, one more. – [Male] I remember yesterday, oh sorry. I remember yesterday you were
talking a little bit about how, what’s it called,
robust it is to degradation and stuff, so I was just
wondering, it made me think of if I have multiple Grok
instances, if you will, existing in the same environment,
and you were to like, if they could I guess make
predictions for the same problem, or be connected in some
way, would certain instances take over certain functions
and others sort of distribute their functionality? – Alright, that’s a tricky
question, and it’s probably deeper than most of the
audience cares about. But, I’ll get into it very quickly. We actually do run multiple instances. Grok is the name of the whole
thing, but we can actually run multiple instances at the same time using these algorithms. And the reason we do that is
because there are some parts of the system which
are not really learned. Just like in your brain, like
your retina is genetically determined, and how it works
is genetically determined. And a different retina,
other animals have different retinas, they may not be as good. So a dog doesn’t see, or
a cat doesn’t see as good as you do in many situations. It has a different retina. So there’s some, there’s
things like that in our system, like learning rates and
how we encode the data, which are sort of more like
biological evolutionary thing. And so, we don’t really know
the best way of encoding the data, we don’t really
know the best learning rates for the system. So we actually run multiples
of these models at the same time and they end up working
to some extent, but some are better than others. And we don’t, you can’t view
them as, not really what you said like they’re learning
different things, but they kind of view the world slightly different. This is like my wife and
I view the world slightly differently, she colors better than I do. I don’t know what the hell
she’s talking about sometimes. But she’s all look at
that color, I’m like what? And she’s an artist. And she’s sensitive to those things. So, different versions of
Grok are like that too. So we’re constantly running
a compilation of these guys and then some are better than others. We kill the ones that don’t do very well. And we produce new ones. So, those were both good questions. I want to get someone
else to ask the question. – [Male] Thank you. – Sure. – [Man] Hi, how does the
performance of your model compare to other machine
learning algorithms? – So the question is how does
the performance of this model compare to other machine
learning algorithms. So, the machine learning
field is very, very large. And, let me just give you,
I’ll answer that question in a couple ways. First of all, you have to ask
yourself what other machine learning models are really
good at time-based data? Very, very few. I can name them on like, two fingers. Then, the next question. How many of them are
machine continuous learning? Online learning? Not batch learning. But, continuous learning. Very, very few. And so you can, what we
claim, what we’re doing here is we’re solving problems
that other machine learning problems don’t solve. It’s not like oh, I can compare
the performance of yours with the compare of these. It’s usually not possible. Now however, I’ll say, that
when we go into a customer, they typically have some solution. It may be sophisticated, it
may be crude, but they have something and they’re
not very happy with it. And so, instead of going in
claiming we don’t go in claiming oh, our algorithm is
better than all the rest. We say no, just give us your
problem, and see if we can improve upon it. ‘Cause usually there aren’t
very good solutions for these type of problems, there
just aren’t good solutions. And so, we come in and we do
much, if we can do much better they’re very, very happy. And so that’s really the claim here. It’s not like, you don’t want
to get there in some sort of, we’re do better at this,
and you do better at this by two percent or something like that. Nah, it’s not about that. It’s about streaming data,
it’s about continuous learning, it’s about making predictions,
you know, and automated, and automated model creation. And there just aren’t
other solutions like that. – [Man] Thanks. – Yeah. – [Male] Hi, I have three questions. – How many? – [Male] What was that? – How many questions? – [Male] Three. – Three! Oh you gotta go for more
than the last guy, okay. – [Male] The first question I was asking, what’s the equivalent of RAM in the brain? – What’s the equivalent of RAM, like RAM– – Random access memory.
– Random access memory? – [Male] Yeah. – Well, did I say there
was an equivalent amount of RAM in the brain? Alright, I don’t even know
how to answer that question. Alright, so, the memory we use, I’ll try. The memory we use in the computer
of RAM is one type, okay. Random access means that
there’s a big list of things you put in an address you
get the results back, right? That’s a type of memory. It is structurally,
completely different than type of memory that I just talked about. There is nothing equivalent
about it what so ever. This is a hierarchical
temporal memory system, and RAM is a linear, flat random access memory. So there’s no equivalent
to that in the brain. Now, I can answer this slightly different. If you thought of RAM in a
computer, typically what’s in the RAM in the computer
is a temporary memory and it’s a state of the system. Alright, like I got the hard
drive is where I’m storing data, and maybe the RAM is the
current state of the system. So if the RAM dies, then
you lose all of the data in the system. If that’s the question you
were asking, and it looks from your face that wasn’t
the question you’re answering, but I’m gonna answer it anyway– – [Male] I’m still
learning about the, yeah. I mean, I’m still learning
about this field, so– – Okay well, let’s put it this way. When our models are running
an instantaneous state, that is like the current
activation, and that’s kinda like what’s kept in a
RAM in a computer memory. But anyway, I’m gonna do the answer. – [Male] But you’re saying
there’s no real equivalent? – There’s no real equivalent. I could have just answered it that way. No, there’s no equivalent. – [Male] My second question
is yeah, I also, is computing model your company’s developing,
is that the only computing model that’s, I mean, I’m
assuming you’re really cutting edge, so– Well, what I was wondering,
what models are your competitors using, I mean, like I said
I’m also kind of studying the companies– – Alright so again, I
don’t want to get too much on the business side of
things, because that’s not what this talk is about, but, we
you know, everyone likes to say this, I don’t know if we
have real competitors. Again, you go in the attributes
of what we’re offering are not available in other systems. Write this down, you got a
piece of paper, you know, continuous learning, temporal
data, automatic model creation okay, and go out and find
companies that do that, okay. So, this is a real good question. A little techy, I didn’t
talk about it in the talk, but I’ll go through it. So the question is, I said
magically here’s these numbers and fields and so on, and
I turn them into Sparse Distributed Representations, right? How do you do that? Let me give you the one
example that’s the simplest one to understand, and then
you can ponder about it. So, imagine I have a number. Energy, price, temperature, right? I got a number, I got
a scale of 45 numbers. It’s on a number line. And I want to turn that
into a Sparse Distributed Representation. So imagine I have this number
line in front of me now, and imagine now I define
you a whole bunch of bits, 200 bits. And the first bit represents zero to 10. And the next bit represents one to 11. And the next bit represents two to 12. And so on. And now, when I have a
number on the number line, say it’s 22, I can go see
which of those bits overlap with that number. And, those are the bits gonna be one. Now this idea came from the cochlear. So we didn’t make this up completely, this is kind of a little
bit how the cochlear works in your ear. And so, if I give you some
number, there will be some bits that come on, because their
bands overlap with that number. And if the number moves
upward a little bit, then one of those bits will turn off
and another bit will turn on. Are you following this? Can you visualize this? So, so we have a Sparse
Distributed Representation, each bit has semantic meaning. I have a number of bits on, it’s sparse. And, similar values have
similar representations. And, that’s how we do it. And it works pretty well. – [Male] When it comes
to semantic categories? – Ah, so when it’s semantic
categories works, you have to understand that everyone in
the system we’re constantly, it’s learning semantic meanings, right? You don’t, you start the
way it works in the brain, and the way it works in our
systems, you start with very low level semantic meanings
like a little bands of frequency, little bands
of numbers, or little patches of visual minds in individual
fields or little things like this. You have to build the more
sophisticated representations as you go. So the very first thing
that happens in our system, is I take all these inputs
from a bunch of fields, these sparse representations,
and I run through something I didn’t talk about today. But basically I have to form
a new Sparse Representation with 2000 bits. My input may not be 2000 bits. My input may be more or less,
and one of two distinguished ones you have to representations
based on the common spacial patterns in the world. So basics says, if I see
coincidences that occur, spacial coincidences that
form representations of them, and then I have sequences
of those, I form– It’s a pretty technical
thing, so I’ll just leave it at that, okay? So the gentleman’s asking
me he say oh, there’s other algorithms like deep belief
networks that you might compare to this, by the way, we
call this cortical learning algorithm, CLA, so you
can compare it to the CLA. And, there’s some, they produce
predictions about biology, about anatomy perhaps, is
that what you’re saying, and then you have the
same type of thing here. Did that cover your question? Okay, first of all, I’ll
disagree with something you said. Deep belief networks are not
nearly as biological grounded as what I presented here. Not even close. The order of the mind stood off. But they are hierarchical in
some sense, and they can deal with time in some sense,
and in that way I like them. They’re in a good direction. But they’re very, very quite different. And they don’t do what I shared here. It’s quite far apart. So they’re not really, easily comparable. None of the customers
we’re working with say hey, I can do this with deep
belief network, they can’t. Now the second part of
your question is whether the biological predictions
are coming out of this. Or can I show behaviors
that we see in biology. And the answer is absolutely, yes. I think presenting that,
we did this original work in with division of
experiments and we’re realizing a lot of physiological
properties that you see, in division system. The answer is absolutely yes. I wasn’t trying to do that
here, I wasn’t showing you that here, but it’s actually very good at that. So I’ll leave it at that. – [Male] What about the entire body? I’m just wondering about emotions. Like, if there’s like
things that are happening in the body, could you put
sensors in there and use that information, or do you do that, or? – Okay, so that’s kind
of a confusing question. But I’ll try to part– – [Male] Well, I’m happy
to try to clarify that– – No, let me try to get it. Let me address the first
few things you said there. Like, what about a body? Can you have sensors in a body? Well, by the way, you do
have sensors in your body. The whole sensory system
called the poke-sensory system which measures the joint
angles and things like this, and we’ll all fit it,
it’s a whole ‘nother thing a lot of people don’t know about it. But, that’s how you
model where your body is. ‘Cause there’s all these
sensors that tell you this. And, we don’t have bodies in our system. Now we’re doing you know,
inference on data streams. And the day that I have a
sensory motor integration loops, and if there was an embodiment
of that, I probably would have bodies in the sensors,
but what kind of body would it be? Is it gonna be humanoid, or
is it gonna be something else? I have a feeling it
would be something else. By the way, this is to speculate here. When I think of sensory motor
integration, most people think of robots. I don’t think about robots. I think about how it is I
would navigate through data. How I would navigate through
a sensory world to know which part of the sensory
stream to pay attention to, and it may not even have
physical movement, the sensory motor integration problem
can be basically dealing with this sort of whole idea of
sort of how do I interact with my world and the cause and effect. Something like that. So that was the last,
probably couple questions, did I miss an important
part of your question? – [Male] Yeah, I mean,
like I said, I guess– – Do you wanna ask a
question about the emotions? Can I really get away without emotions? – [Male] Yeah, I mean basically
could there be intelligence in the emotions? – Yeah, I think you can have it. But as I define it, if you want
to define it as human-like, then no. If you want to be human-like, then you have to have emotions. If you want to pass the Turing
Test, you’re gonna have to have emotions. But to do the things that I
talked about, make the world better and explore the
universe, I don’t think so. – [Male] Not necessary. – Not necessary. All you would– In the end, what emotions
do, from the biological point of view, they basically,
there’s sort of a switch to tell the neocortex to
learn this very rapidly. So when an emotional thing
happens, the amygdala will say you know, remember this! And store it very quickly. And that’s how it kind of
plays with the neocortex. We can do things like that. But, it’s not an essential
component of what intelligence is, and by the way, you
mentioned the music thing. You said you’re into music. I’ll mention, there’s a guy
Charlie Gillion, and I was, who plays with Counting Crows,
and he’s one of the musicians there, and this guy is a
rock star computer machinery geek, you wouldn’t know this. But when he’s not turning,
this is what he does. And, he’s talked about coming
and visit us, because he’s trying to build, I don’t
want to give too much away, but he’s trying to build
a product in the world of machine generated music. And he’s trying to
understand how you do this, and he’s really intrigued
by the kind of algorithms we have, so we’ll see if
something comes out of that. Alright, one more person
with questions, this woman who’s up next. You’re the last one, so make it great! Make it last! – Hi, well, it’s probably
like the simplest question of all. Actually, was like really
similar to what he asked. But, I was going to ask
about, so I was going to ask about emotions if you felt
that the reason you disregarded it is because it’s just
so hard to have an emotion in like, an artificial
intelligence, or because you just felt that that actually
hindered intelligence. So– – It’s the latter. Actually, from a neuromechanism
point of view, emotion systems in your brain are fairly small. I’m not sure we understand them. I don’t study them exactly. I understand how they
interact with the neocortex, which is fairly simple. And so I think I have a
big feeling about that. But, it’s more, it’s more
a matter of to what purpose do you want emotions? What’s the purp– Why do you want to do this? If there was a need for
it, well, we’ll do it! But, I don’t see any
reason why you can’t do it. But, but, the only reason
I didn’t put it on my list is I don’t think it’s
an essential ingredient. I don’t see situations that
when you absolutely have to have emotions to do the things I
talked about: streaming data, build models of the world,
understand the structure of the world, make
predictions about the world. Those do not require emotions. Emotions are more for
like, what was important or what wasn’t important? How do I prioritize
between various things, and things like that. So, and I don’t think
there’s anything mystical about emotions, there’s
nothing like oh you can’t model that, you know, in the end
I mentioned this in my talk yesterday, I didn’t bring
it up today, but you know, the brain is just a bunch of neurons. There’s nothing else. Everything is just a bunch of neurons. And, you know, there’s no
other magic going on up there. So, if you can understand
how those neurons work and how they play together in
the way I talked about then, yeah, you could model that too. – [Woman] So then, there’s
nothing human that can be in a machine? – Is there nothing human
that can’t be in a machine? Again, I said that I didn’t
think that was the goal, and I don’t think that’s gonna happen. Because I don’t think
there’s no reason to do that. But on a theoretical level,
I don’t see anything that’s in a human that couldn’t
be, and again, you’re saying a machine, don’t think
of it like some you know, mechanics and a machine. It’s a very complex memory
system, and yeah, I don’t think there’s anything magic going on there. – [Woman] May I ask one more question? – You don’t like my
answer, that’s why you ask another question. (audience laughing) – [Woman] No, it’s different, actually. It’s on the side, but, kind of– – I’m gonna have to wrap
up soon, I only have one person so real quick– – [Woman] Okay, so basically
you said we have too much data and it’s like, building up. And, do you propose a solution for that? – Yeah, in some sense. This is switching away from
brains and neuroscience. The solution is that we are
not gonna save all the data in the world. There’s no point in saving it. Your brain doesn’t save all
the sensory information, there’s no point in saving
second by second energy information from a billion buildings. The point, the whole way of
getting around that problem is to handle the data
immediately, feed it into models, billions of models, have
them act on it immediately, and get rid of it. That’s the solution to big data. Okay, let’s the other woman,
do you want to ask a question? Okay, can I have one more question. I lied, I’ll let one more question in. – [Female] I’m just curious
that like when do you think the fluid robot really exist? – When would the fluid
robotics really exist? – [Woman] When do you think, yeah. – Boy, that’s a tough question. – [Woman] Can I see that before I die? – Will you see it before you die? How old are you? – [Woman] I’m 21. – 31? I think so, yeah. I’m not sure I’ll see it before I die. Thank you very much, good luck. (audience clapping) (soft string music)