[Music]
[Applause]
[Music]
[Applause]
so artificial intelligence have recently
gained a lot of attention right with a
little bit of acceleration I think I
came here hey I am machine nonny you
know thousand times per day in the news
in the newspapers or TBS and so on as
you all know recently there was a
historical event where machine actually
peed our human champion in the coal game
which has been considered very hard to
do
there was also news of newspaper
articles about the machine that can
diagnose the cancer even with the
slightest hint in the image of the
medical image right and there was also
there is also soft here that can trade
the stock and make money out of it
indeed AI is in your computers and
smartphone as well
whenever you browse the website some of
the F site actually try to predict your
interest and come up with some product
that you might be interested in buying
right in all of this area and other area
areas including the autonomous speaker
the AI is now really making a huge
impact in our life so I also want to
bring your attention to another exciting
technology which is called Internet of
Things or IOT the essential idea of IOT
is to attach a small computer onto
everyday object and by doing so you can
include the object into the traditional
Internet with this capability what you
can do is you can actually
monitor control and observe a lot of
things for example you can see the
temperature we reached refrigerator and
also change it you can turn it on turn
it off your light curve in the house
while you are traveling you can also
make sure that there’s no water leak in
your water pass it in the sink ai and
IOT what would be the answer I believe
the answer is smart things in fact we
already seeing smart things in our life
the best very good example is the smart
speaker this speaker can detect human
voice in the noisy environment and do
something for you detecting a human
voice in a noisy environment sounds like
a very simple task to us as many human
however for machine to do that the
machine must acquire a knowledge or an
idea what makes human voice different
from all the different kinds of noise
this knowledge is technically called a
model or more specifically in this case
model that detect human voice or human
voice detection model and the process of
training this model making this model is
called machine learning our research
team at Columbia University have
recently developed very low power boost
detection system and that system has
this voice detection model and you can
see that the graphical representation of
the model in the screen essentially the
model is collection of mathematical
formula that you want to evaluate again
and again with a new input
see whether there’s a voice or not in
the ambient sound there’s a lot of
exciting application in smart things
when one up to my favorite is actually
smart factory which consists of a lot of
different smart things for example in
the smart factory you can imagine a law
but they can fetch the parts that you
need upon your voice command you can
also consider a poor such as welding
machine they can not only do the welding
but also monitor the quality of the
welding real-time and report the quality
of the job to the users or the operator
or a supervisor to make sure that
welding is perfectly done all the time
this type of puzzle capability improves
the productivity of the factory while
reducing the cost of operating the
factory so let me switch your gear a
little bit here and tell you a little
bit about how we can actually make the
smart things today
first what we’re going to do is we’re
going to use a very powerful cloud
computer to build a model and this model
will be deployed onto the small mobile
device like your smartphone and finally
with this model your smart phone can
detect your voice and other intelligent
things to do a intelligent task is
actually not simple for example if I
want to make my smart phone to detect a
specific keyword let’s say happy or any
other keyword that you can imagine it
actually requires to compute roughly
five million multiplications and five
million additions moreover you have to
do it every 100 millisecond so this is
non-trivial workload even put a very
high standard of the
smartphone that we develop and I just
told you that the mother is all trained
in the cloud computer right however it
would be even better if a mobile device
can train the model in field locally
using the or interesting private data so
like this and after that all these
mother’s trained locally in multiples of
these mobile devices can be unloaded
unloaded to the cloud computer and cloud
computers combine all of them and make
the create model and apparently you want
to download this model to the order
smart devices so that they can do job
more intelligently and more accurately a
training is not that easy ding let’s say
I want to make a machine that detect a
frog in the picture right that task
lovely involves analyzed or one picture
however if I want to build a model that
detect the frog that requires tunnel
eyes let’s say about 50,000 images over
hundred times eight order five million
image so there’s no orders of magnitude
difference between post training
intelligent work versus training so in
order to really enable these smart
things we must address this
computational complexity problem so
let’s go a little bit deeper and see how
computer computers actually do the
computation let’s consider this
implication a plus B is the C right so
first step computer actually do this in
two steps the first step the computer
must do is data movement
what that means is they grab a and B
from the memory and move all the way to
the input of the another hardware called
computers actually perform addition and
produced C so that’s the whole story
right but the problem is actually in
modern computer systems moving data is
way more expensive and actually do
computation this is because the memory
and the computational elements are
physically separated transmitting the
information over a distance whether it’s
micrometers to millimeters or centimeter
takes time and energy so this problem
called memory word problem has been
identified about 10 years ago by
Professor wolf and McKee and what they
observe is the speed of the processor
increases much faster than the time to
get the data from the memory and this
difference for now actually increase the
difference of the speed by lawfully 100
times 2,000 times so memory word problem
is very fundamental problem in today’s
computing system any computers that you
have laptop smartphone PC servers
anything so this is the very fundamental
problem that we need to solve so
basically we want to reduce this data
movement between memories and our smell
unit so there’s a lotta two different
approaches in my opinion the first
approach is to use a small memory in the
arithmetic unit indeed there is a small
memory embedded in the arithmetic
hardware that key challenge is this
memory is not large enough to fit all
the necessary data that you want to use
so the key design goal opie to make
sure there use this embedded memory
maximally so that we can reduce the data
movement to be outside large memory so
toward this core our team at Columbia
University recently develop a new tip
learning algorithm called liquor C
binary linear in the talk so this
learning algorithm go through the mainly
three step first training second
knowledge consolidation so I want to
compress the knowledge and after that I
want to use the space that I just make
and I want to repeat this multiple times
until I lead you to the target accuracy
of the model so this model maximally
used the embedded memory that is
embedded in the erase memory unit so you
can reduce the data movement to all from
the large outside of the memory and this
basically translated huge energy savings
and the speed-up of the training process
the other approach that we can consider
is somewhat similar to previous approach
but are we different so what we can do
is basically this so-called in-memory
computing so what that means is instead
of transmitting all the raw data we can
actually do a little bit of computation
inside of the memory so compress the
data and pass it to the arithmetic unit
so that we can reduce the data movement
to do so we have to add a small hardware
in the memory that can do a little bit
of arithmetic operation and this must be
chosen very carefully to support your
preferred algorithm very la in a big
time
so in this core our team in the
collaboration with the original State
University develop and prototype a new
memory hardware they can actually do
compute inside it it the name is called
X Norris Ram
and we add a small computational element
they can do this so forth multiply and
accumulate operation and multiply and
accumulate operation is the fundamental
very basic massively used computational
corner in deep learnings as well as the
demurrer Network so having this memory
hardware we can actually transmit much a
smaller amount of data because we
already compute the value and the result
will be pass to the arithmetic unit and
this again translated about two orders
of magnitude improvement in terms of the
so I would like to conclude at this
moment thank you for your listening my
talk the future of AI is actually not
small it’s bright and one of my what
needs to be small the AI size needs to
be small so that we can fit it in the
smart things so that’s all this is Ming
Gustav from Columbia University thank
you very much [Applause]
[Applause]
[Music]
[Applause]
so artificial intelligence have recently
gained a lot of attention right with a
little bit of acceleration I think I
came here hey I am machine nonny you
know thousand times per day in the news
in the newspapers or TBS and so on as
you all know recently there was a
historical event where machine actually
peed our human champion in the coal game
which has been considered very hard to
do
there was also news of newspaper
articles about the machine that can
diagnose the cancer even with the
slightest hint in the image of the
medical image right and there was also
there is also soft here that can trade
the stock and make money out of it
indeed AI is in your computers and
smartphone as well
whenever you browse the website some of
the F site actually try to predict your
interest and come up with some product
that you might be interested in buying
right in all of this area and other area
areas including the autonomous speaker
the AI is now really making a huge
impact in our life so I also want to
bring your attention to another exciting
technology which is called Internet of
Things or IOT the essential idea of IOT
is to attach a small computer onto
everyday object and by doing so you can
include the object into the traditional
Internet with this capability what you
can do is you can actually
monitor control and observe a lot of
things for example you can see the
temperature we reached refrigerator and
also change it you can turn it on turn
it off your light curve in the house
while you are traveling you can also
make sure that there’s no water leak in
your water pass it in the sink ai and
IOT what would be the answer I believe
the answer is smart things in fact we
already seeing smart things in our life
the best very good example is the smart
speaker this speaker can detect human
voice in the noisy environment and do
something for you detecting a human
voice in a noisy environment sounds like
a very simple task to us as many human
however for machine to do that the
machine must acquire a knowledge or an
idea what makes human voice different
from all the different kinds of noise
this knowledge is technically called a
model or more specifically in this case
model that detect human voice or human
voice detection model and the process of
training this model making this model is
called machine learning our research
team at Columbia University have
recently developed very low power boost
detection system and that system has
this voice detection model and you can
see that the graphical representation of
the model in the screen essentially the
model is collection of mathematical
formula that you want to evaluate again
and again with a new input
see whether there’s a voice or not in
the ambient sound there’s a lot of
exciting application in smart things
when one up to my favorite is actually
smart factory which consists of a lot of
different smart things for example in
the smart factory you can imagine a law
but they can fetch the parts that you
need upon your voice command you can
also consider a poor such as welding
machine they can not only do the welding
but also monitor the quality of the
welding real-time and report the quality
of the job to the users or the operator
or a supervisor to make sure that
welding is perfectly done all the time
this type of puzzle capability improves
the productivity of the factory while
reducing the cost of operating the
factory so let me switch your gear a
little bit here and tell you a little
bit about how we can actually make the
smart things today
first what we’re going to do is we’re
going to use a very powerful cloud
computer to build a model and this model
will be deployed onto the small mobile
device like your smartphone and finally
with this model your smart phone can
detect your voice and other intelligent
things to do a intelligent task is
actually not simple for example if I
want to make my smart phone to detect a
specific keyword let’s say happy or any
other keyword that you can imagine it
actually requires to compute roughly
five million multiplications and five
million additions moreover you have to
do it every 100 millisecond so this is
non-trivial workload even put a very
high standard of the
smartphone that we develop and I just
told you that the mother is all trained
in the cloud computer right however it
would be even better if a mobile device
can train the model in field locally
using the or interesting private data so
like this and after that all these
mother’s trained locally in multiples of
these mobile devices can be unloaded
unloaded to the cloud computer and cloud
computers combine all of them and make
the create model and apparently you want
to download this model to the order
smart devices so that they can do job
more intelligently and more accurately a
training is not that easy ding let’s say
I want to make a machine that detect a
frog in the picture right that task
lovely involves analyzed or one picture
however if I want to build a model that
detect the frog that requires tunnel
eyes let’s say about 50,000 images over
hundred times eight order five million
image so there’s no orders of magnitude
difference between post training
intelligent work versus training so in
order to really enable these smart
things we must address this
computational complexity problem so
let’s go a little bit deeper and see how
computer computers actually do the
computation let’s consider this
implication a plus B is the C right so
first step computer actually do this in
two steps the first step the computer
must do is data movement
what that means is they grab a and B
from the memory and move all the way to
the input of the another hardware called
computers actually perform addition and
produced C so that’s the whole story
right but the problem is actually in
modern computer systems moving data is
way more expensive and actually do
computation this is because the memory
and the computational elements are
physically separated transmitting the
information over a distance whether it’s
micrometers to millimeters or centimeter
takes time and energy so this problem
called memory word problem has been
identified about 10 years ago by
Professor wolf and McKee and what they
observe is the speed of the processor
increases much faster than the time to
get the data from the memory and this
difference for now actually increase the
difference of the speed by lawfully 100
times 2,000 times so memory word problem
is very fundamental problem in today’s
computing system any computers that you
have laptop smartphone PC servers
anything so this is the very fundamental
problem that we need to solve so
basically we want to reduce this data
movement between memories and our smell
unit so there’s a lotta two different
approaches in my opinion the first
approach is to use a small memory in the
arithmetic unit indeed there is a small
memory embedded in the arithmetic
hardware that key challenge is this
memory is not large enough to fit all
the necessary data that you want to use
so the key design goal opie to make
sure there use this embedded memory
maximally so that we can reduce the data
movement to be outside large memory so
toward this core our team at Columbia
University recently develop a new tip
learning algorithm called liquor C
binary linear in the talk so this
learning algorithm go through the mainly
three step first training second
knowledge consolidation so I want to
compress the knowledge and after that I
want to use the space that I just make
and I want to repeat this multiple times
until I lead you to the target accuracy
of the model so this model maximally
used the embedded memory that is
embedded in the erase memory unit so you
can reduce the data movement to all from
the large outside of the memory and this
basically translated huge energy savings
and the speed-up of the training process
the other approach that we can consider
is somewhat similar to previous approach
but are we different so what we can do
is basically this so-called in-memory
computing so what that means is instead
of transmitting all the raw data we can
actually do a little bit of computation
inside of the memory so compress the
data and pass it to the arithmetic unit
so that we can reduce the data movement
to do so we have to add a small hardware
in the memory that can do a little bit
of arithmetic operation and this must be
chosen very carefully to support your
preferred algorithm very la in a big
time
so in this core our team in the
collaboration with the original State
University develop and prototype a new
memory hardware they can actually do
compute inside it it the name is called
X Norris Ram
and we add a small computational element
they can do this so forth multiply and
accumulate operation and multiply and
accumulate operation is the fundamental
very basic massively used computational
corner in deep learnings as well as the
demurrer Network so having this memory
hardware we can actually transmit much a
smaller amount of data because we
already compute the value and the result
will be pass to the arithmetic unit and
this again translated about two orders
of magnitude improvement in terms of the
so I would like to conclude at this
moment thank you for your listening my
talk the future of AI is actually not
small it’s bright and one of my what
needs to be small the AI size needs to
be small so that we can fit it in the
smart things so that’s all this is Ming
Gustav from Columbia University thank
you very much [Applause]
Please follow and like us: