
[Music]
Computers today can do incredible things
they allow us to browse the internet
send emails and most importantly share
memes on Facebook they also enable
life-saving medical imaging advanced
scientific simulations and global
financial networks without question
computers have revolutionized society
but what if we wanted to build a
computer that could see in here and
perceive the world or a robot that could
act with the agility and intelligence of
a cat or dog this is of course the goal
of artificial intelligence but if we
could build such a computer what would
it tell us about ourselves this is the
question that inspired me to become a
researcher in artificial intelligence as
far as we can tell we are our brains all
of our thoughts memories personalities
and motivations are encoded up there
somehow understanding the brain would
open up new doors in psychology
sociology and mental health one way to
get a better understanding of the brain
is to try to build one so why can’t we
put simply brains in computers were
built to solve two very different sets
of problems computers were built to
follow instructions executing one
calculation after another very rapidly
brains were built to survive using
intuition and continuously adapting to a
changing environment intuition is the
ability to know the answer to something
without needing to know why it’s sort of
our gut instinct for instance when you
see your friend on the street you
immediately recognize him or her you
don’t need to stop and think what color
is this person’s hair are they wearing
glasses you just know it’s them all at
once adaptation is the ability to change
in response to your environment
learning for instance is a form of
adaptation we change our understanding
of the world by interacting with it it
turns out that adaptation and intuition
are really difficult to program in the
conventional sense and that’s because
the real world is complicated and you
can’t possibly include every scenario
that a computer might run into but over
the past five years or so we have begun
making incredible progress in
replicating these facets of intelligence
I’m sure you’ve heard of self-driving
cars or notice that automatic language
translation has improved dramatically or
maybe you’ve noticed that now you can
speak to your smartphone these are all
because of breakthrough advancements in
the science of artificial neural
networks neural networks are algorithms
that are inspired by biological brains
they were originally designed to help us
understand how the neurons in their
connections or synapses in our own
brains work since they were first
invented neural networks have come a
long way and we can now do pretty
incredible things with them for instance
we can teach them to translate between
two languages or drive a car or
recognize cats in images these networks
are what’s known as massively parallel
they allow a lot of simple computations
to be done at the same time but our
current processors don’t work like this
they process one piece of information
after another this makes our current
processors really bad at simulating
neural networks
despite this neural networks have become
state-of-the-art algorithms in many
fields but even our largest most
powerful networks are still only a tiny
fraction of the size of even a mouse’s
brain even these comparatively small
networks can take months to train on our
if we is we try to scale the network by
adding more and more neurons the amount
of time that it takes to simulate them
takes longer and longer eventually we
reach a point where it’s simply not
practical to simulate for practical
neural networks such as in speech and
video recognition this point is about a
million neurons or the size of a
honeybees brain a mouse’s brain let
alone in humans is beyond reach for now
if we want to build truly scalable
neural networks we need to fundamentally
rethink our computing hardware in
particular we need to design processors
specifically for neural networks with a
massively parallel architecture this
would allow us to run and train our
neural networks in real time independent
of the size of the network the field
tackling this problem is called
neuromorphic computing neuromorphic
computing moves beyond the standard
computing paradigm and builds physical
neuron and synapse circuits onto a chip
mimicking biological brains my hope is
that these neuromorphic processors will
someday help us understand our own
unfortunately these neuromorphic
processors are not scalable either but
in a different way whereas conventional
processors take too much time to
simulate neuromorphic architectures take
up too much space as you try to pack
more neurons onto a chip the amount of
wiring that you need to connect all of
them together quickly gets out of hand
even for relatively small networks not
coincidentally this is a constraint
that’s also shared by biological brains
in both neuromorphic architectures and
biological brains you need to connect
large numbers of neurons with modifiable
connections in a scalable way the
difference between the two is that
biological brains have found a solution
in real brains we don’t see these fully
connected grid like networks of neurons
like we commonly see in neuromorphic
architectures
instead they’re sparsely connected where
two neurons that are very closely very
close together have a high probability
of being connected while neurons that
are very far apart have a much lower but
still significant probability of being
connected this creates a network that’s
very well connected while only requiring
a tiny fraction of the total possible
number of connections fewer connections
means less wiring and less wiring means
a more scalable architecture this type
of sparse random connectivity is known
as small world connectivity small world
nests embodies the concept of six
degrees of separation the idea that you
are connected to any other person on
earth through at most six people
similarly every neuron in your brain is
connected to any other neuron through a
very small number of intermediate
neurons fully parallel small world
connectivity is really difficult to
achieve in hardware but being able to
replicate it would not only allow for
scaleable neuromorphic architectures it
would also give neuroscientists a
valuable new tool for building large
scale models of biological brains this
is the problem at the center of my
research at the University of Florida we
have created a new method for connecting
large arrays of artificial silicon
neurons using ultra-thin synaptic wires
that can connect neurons with realistic
synapses the resulting network is sparse
and random with small world’s
connectivity very similar to the types
of connectivity that we see in
biological brains the fact that these
neuromorphic architectures require
similar patterns of connectivity as
biological brains in order to scale
suggests that this is their purpose in
we also know that the algorithms that we
use to train our artificial neural
networks are incompatible with the
physical structure of the brain it’s
simply not physically possible for the
brain to be implementing these
algorithms in their current form so in
order to understand the hardware or the
software of the brain the algorithms we
also need to understand its hardware
this means creating artificial neural
networks with the same physical
constraints as biological brains
overcoming these constraints such as the
scaling of wiring complexity can inform
us as to how and why the brain chooses
certain designs over others I believe
that this process of implementing in
overcoming the physical constraints of
the brain in Hardware will allow us to
establish a common language between
artificial intelligence and neuroscience
fields that are largely disjoint today
this common language would allow a huge
amount of information to flow between AI
and neuroscience and help us come to a
better understanding of the brain
obviously understanding the brain is a
huge challenge people have been trying
to understand the brain ever since we
realized that the brain is the seat of
consciousness but along the way we have
accumulated a huge amount of knowledge
and we’re now approaching a fundamental
understanding of how neural networks
function as evidenced by the incredible
things that we can do with them when I
first started my research the brain
seemed like a total mystery it seemed
completely incomprehensible but the more
I study it the more I’m convinced that
it can be understood the pace of
progress in neural networks right now is
breathtaking
I really encourage anyone who’s
interested in AI or neuroscience to just
jump right in there has never been a
better time to make an impact thank you you
Computers today can do incredible things
they allow us to browse the internet
send emails and most importantly share
memes on Facebook they also enable
life-saving medical imaging advanced
scientific simulations and global
financial networks without question
computers have revolutionized society
but what if we wanted to build a
computer that could see in here and
perceive the world or a robot that could
act with the agility and intelligence of
a cat or dog this is of course the goal
of artificial intelligence but if we
could build such a computer what would
it tell us about ourselves this is the
question that inspired me to become a
researcher in artificial intelligence as
far as we can tell we are our brains all
of our thoughts memories personalities
and motivations are encoded up there
somehow understanding the brain would
open up new doors in psychology
sociology and mental health one way to
get a better understanding of the brain
is to try to build one so why can’t we
put simply brains in computers were
built to solve two very different sets
of problems computers were built to
follow instructions executing one
calculation after another very rapidly
brains were built to survive using
intuition and continuously adapting to a
changing environment intuition is the
ability to know the answer to something
without needing to know why it’s sort of
our gut instinct for instance when you
see your friend on the street you
immediately recognize him or her you
don’t need to stop and think what color
is this person’s hair are they wearing
glasses you just know it’s them all at
once adaptation is the ability to change
in response to your environment
learning for instance is a form of
adaptation we change our understanding
of the world by interacting with it it
turns out that adaptation and intuition
are really difficult to program in the
conventional sense and that’s because
the real world is complicated and you
can’t possibly include every scenario
that a computer might run into but over
the past five years or so we have begun
making incredible progress in
replicating these facets of intelligence
I’m sure you’ve heard of self-driving
cars or notice that automatic language
translation has improved dramatically or
maybe you’ve noticed that now you can
speak to your smartphone these are all
because of breakthrough advancements in
the science of artificial neural
networks neural networks are algorithms
that are inspired by biological brains
they were originally designed to help us
understand how the neurons in their
connections or synapses in our own
brains work since they were first
invented neural networks have come a
long way and we can now do pretty
incredible things with them for instance
we can teach them to translate between
two languages or drive a car or
recognize cats in images these networks
are what’s known as massively parallel
they allow a lot of simple computations
to be done at the same time but our
current processors don’t work like this
they process one piece of information
after another this makes our current
processors really bad at simulating
neural networks
despite this neural networks have become
state-of-the-art algorithms in many
fields but even our largest most
powerful networks are still only a tiny
fraction of the size of even a mouse’s
brain even these comparatively small
networks can take months to train on our
if we is we try to scale the network by
adding more and more neurons the amount
of time that it takes to simulate them
takes longer and longer eventually we
reach a point where it’s simply not
practical to simulate for practical
neural networks such as in speech and
video recognition this point is about a
million neurons or the size of a
honeybees brain a mouse’s brain let
alone in humans is beyond reach for now
if we want to build truly scalable
neural networks we need to fundamentally
rethink our computing hardware in
particular we need to design processors
specifically for neural networks with a
massively parallel architecture this
would allow us to run and train our
neural networks in real time independent
of the size of the network the field
tackling this problem is called
neuromorphic computing neuromorphic
computing moves beyond the standard
computing paradigm and builds physical
neuron and synapse circuits onto a chip
mimicking biological brains my hope is
that these neuromorphic processors will
someday help us understand our own
unfortunately these neuromorphic
processors are not scalable either but
in a different way whereas conventional
processors take too much time to
simulate neuromorphic architectures take
up too much space as you try to pack
more neurons onto a chip the amount of
wiring that you need to connect all of
them together quickly gets out of hand
even for relatively small networks not
coincidentally this is a constraint
that’s also shared by biological brains
in both neuromorphic architectures and
biological brains you need to connect
large numbers of neurons with modifiable
connections in a scalable way the
difference between the two is that
biological brains have found a solution
in real brains we don’t see these fully
connected grid like networks of neurons
like we commonly see in neuromorphic
architectures
instead they’re sparsely connected where
two neurons that are very closely very
close together have a high probability
of being connected while neurons that
are very far apart have a much lower but
still significant probability of being
connected this creates a network that’s
very well connected while only requiring
a tiny fraction of the total possible
number of connections fewer connections
means less wiring and less wiring means
a more scalable architecture this type
of sparse random connectivity is known
as small world connectivity small world
nests embodies the concept of six
degrees of separation the idea that you
are connected to any other person on
earth through at most six people
similarly every neuron in your brain is
connected to any other neuron through a
very small number of intermediate
neurons fully parallel small world
connectivity is really difficult to
achieve in hardware but being able to
replicate it would not only allow for
scaleable neuromorphic architectures it
would also give neuroscientists a
valuable new tool for building large
scale models of biological brains this
is the problem at the center of my
research at the University of Florida we
have created a new method for connecting
large arrays of artificial silicon
neurons using ultra-thin synaptic wires
that can connect neurons with realistic
synapses the resulting network is sparse
and random with small world’s
connectivity very similar to the types
of connectivity that we see in
biological brains the fact that these
neuromorphic architectures require
similar patterns of connectivity as
biological brains in order to scale
suggests that this is their purpose in
we also know that the algorithms that we
use to train our artificial neural
networks are incompatible with the
physical structure of the brain it’s
simply not physically possible for the
brain to be implementing these
algorithms in their current form so in
order to understand the hardware or the
software of the brain the algorithms we
also need to understand its hardware
this means creating artificial neural
networks with the same physical
constraints as biological brains
overcoming these constraints such as the
scaling of wiring complexity can inform
us as to how and why the brain chooses
certain designs over others I believe
that this process of implementing in
overcoming the physical constraints of
the brain in Hardware will allow us to
establish a common language between
artificial intelligence and neuroscience
fields that are largely disjoint today
this common language would allow a huge
amount of information to flow between AI
and neuroscience and help us come to a
better understanding of the brain
obviously understanding the brain is a
huge challenge people have been trying
to understand the brain ever since we
realized that the brain is the seat of
consciousness but along the way we have
accumulated a huge amount of knowledge
and we’re now approaching a fundamental
understanding of how neural networks
function as evidenced by the incredible
things that we can do with them when I
first started my research the brain
seemed like a total mystery it seemed
completely incomprehensible but the more
I study it the more I’m convinced that
it can be understood the pace of
progress in neural networks right now is
breathtaking
I really encourage anyone who’s
interested in AI or neuroscience to just
jump right in there has never been a
better time to make an impact thank you you
Please follow and like us: