I’m super excited to talk to you today about the present and future of
about the present and future of
about the present and future of artificial intelligence whenever there’s
artificial intelligence whenever there’s
artificial intelligence whenever there’s a buzzword and a complex subject matter
a buzzword and a complex subject matter
a buzzword and a complex subject matter it’s usually good to start with a
it’s usually good to start with a
it’s usually good to start with a definition but it’s actually a little
definition but it’s actually a little
definition but it’s actually a little tricky because the definition of
tricky because the definition of
tricky because the definition of artificial intelligence seems to be
artificial intelligence seems to be
artificial intelligence seems to be constantly moving whenever we solve a
constantly moving whenever we solve a
constantly moving whenever we solve a problem we don’t quite call it
problem we don’t quite call it
problem we don’t quite call it artificial intelligence anymore
artificial intelligence anymore
artificial intelligence anymore it started with chess a lot of smart
it started with chess a lot of smart
it started with chess a lot of smart researchers looked at other smart people
researchers looked at other smart people
researchers looked at other smart people and thought well we’re really good at
and thought well we’re really good at
and thought well we’re really good at math and logic and playing complex games
math and logic and playing complex games
math and logic and playing complex games like chess and so they started working
like chess and so they started working
like chess and so they started working on those kinds of problems thinking that
on those kinds of problems thinking that
on those kinds of problems thinking that once they solve them a lot of other
once they solve them a lot of other
once they solve them a lot of other things will just fall into place but it
things will just fall into place but it
things will just fall into place but it didn’t quite because those were
didn’t quite because those were
didn’t quite because those were simulated environments that didn’t have
simulated environments that didn’t have
simulated environments that didn’t have the right to the same kind of noise that
the right to the same kind of noise that
the right to the same kind of noise that we have in the real world
we have in the real world
we have in the real world so now research has actually shifted
so now research has actually shifted
so now research has actually shifted largely from playing games which is
largely from playing games which is
largely from playing games which is still an important area and can feature
still an important area and can feature
still an important area and can feature some things – things that we didn’t used
some things – things that we didn’t used
some things – things that we didn’t used to consider as that much of high
to consider as that much of high
to consider as that much of high intelligence just understanding spoken
intelligence just understanding spoken
intelligence just understanding spoken words seems relatively simple we can all
words seems relatively simple we can all
words seems relatively simple we can all do it but that was actually a really
do it but that was actually a really
do it but that was actually a really hard problem up until 2010 when deep
hard problem up until 2010 when deep
hard problem up until 2010 when deep learning changed it and is able to make
learning changed it and is able to make
learning changed it and is able to make much more progress on this and now we
much more progress on this and now we
much more progress on this and now we don’t call it AI anymore it’s just Siri
don’t call it AI anymore it’s just Siri
don’t call it AI anymore it’s just Siri it’s just a speech recognition software
it’s just a speech recognition software
it’s just a speech recognition software but that was a really hard problem that
but that was a really hard problem that
but that was a really hard problem that we weren’t able to solve and there’s
we weren’t able to solve and there’s
we weren’t able to solve and there’s still some tricky issues in research in
still some tricky issues in research in
still some tricky issues in research in it another area that deep learning has
it another area that deep learning has
it another area that deep learning has made a huge amount of progress in in
made a huge amount of progress in in
made a huge amount of progress in in recent years is computer vision namely
recent years is computer vision namely
recent years is computer vision namely image classification and this is the one
image classification and this is the one
image classification and this is the one time I’ll try to explain a little bit to
time I’ll try to explain a little bit to
time I’ll try to explain a little bit to you what these kinds of models do and
you what these kinds of models do and
you what these kinds of models do and how they work one of the most important
how they work one of the most important
how they work one of the most important ideas of recent years in the AI is to
ideas of recent years in the AI is to
ideas of recent years in the AI is to have so-called end-to-end trainable
have so-called end-to-end trainable
have so-called end-to-end trainable models where we take in raw input for
models where we take in raw input for
models where we take in raw input for instance the pixels of an image and want
instance the pixels of an image and want
instance the pixels of an image and want to predict a final output for instance
to predict a final output for instance
to predict a final output for instance is there a cat or a dog or house or
is there a cat or a dog or house or
is there a cat or a dog or house or clock in that image and so as we put
clock in that image and so as we put
clock in that image and so as we put that raw input the pixels into these
that raw input the pixels into these
that raw input the pixels into these models they keep trying to learn more
models they keep trying to learn more
models they keep trying to learn more and more complex representations so as
and more complex representations so as
and more complex representations so as they
they
they start looking at the pixels the first
start looking at the pixels the first
start looking at the pixels the first layer might only identify simple edges
layer might only identify simple edges
layer might only identify simple edges and blobs which actually turns out to
and blobs which actually turns out to
and blobs which actually turns out to also have good correlation to the early
also have good correlation to the early
also have good correlation to the early visual cortex in the human brain but
visual cortex in the human brain but
visual cortex in the human brain but then as they go to the next layer they
then as they go to the next layer they
then as they go to the next layer they combine these blobs and colors and edges
combine these blobs and colors and edges
combine these blobs and colors and edges to more complex textures and then as
to more complex textures and then as
to more complex textures and then as they go further and deeper into these
they go further and deeper into these
they go further and deeper into these different layers they’ll identify object
different layers they’ll identify object
different layers they’ll identify object parts and eventually combine those
parts and eventually combine those
parts and eventually combine those object parts to identify full objects
object parts to identify full objects
object parts to identify full objects and that was really really hard and
and that was really really hard and
and that was really really hard and initially people tried to manually
initially people tried to manually
initially people tried to manually identify oh if there’s a cat then maybe
identify oh if there’s a cat then maybe
identify oh if there’s a cat then maybe their whiskers here then this might
their whiskers here then this might
their whiskers here then this might improve to increase the probability of a
improve to increase the probability of a
improve to increase the probability of a cat and things like that and now this
cat and things like that and now this
cat and things like that and now this entire process all these visualizations
entire process all these visualizations
entire process all these visualizations that you see here on this slide they’re
that you see here on this slide they’re
that you see here on this slide they’re all learned automatically just by giving
all learned automatically just by giving
all learned automatically just by giving it a lot of supervised data here’s an
it a lot of supervised data here’s an
it a lot of supervised data here’s an image and it’s pixels here’s the output
image and it’s pixels here’s the output
image and it’s pixels here’s the output that we care about
that we care about
that we care about now we’ve actually been able to combine
now we’ve actually been able to combine
now we’ve actually been able to combine in computer vision even with some
in computer vision even with some
in computer vision even with some language processing and we can do quite
language processing and we can do quite
language processing and we can do quite amazing things in the last couple of
amazing things in the last couple of
amazing things in the last couple of years here you see a visualization of a
years here you see a visualization of a
years here you see a visualization of a recent paper of collaborators of mine
recent paper of collaborators of mine
recent paper of collaborators of mine where we color code where the algorithm
where we color code where the algorithm
where we color code where the algorithm is paying attention to as it’s trying to
is paying attention to as it’s trying to
is paying attention to as it’s trying to generate a description of an image so
generate a description of an image so
generate a description of an image so you have a little girl sitting on a
you have a little girl sitting on a
you have a little girl sitting on a bench holding an umbrella and you see
bench holding an umbrella and you see
bench holding an umbrella and you see that indeed when it’s generating the
that indeed when it’s generating the
that indeed when it’s generating the word girl it is looking at the girl when
word girl it is looking at the girl when
word girl it is looking at the girl when it’s generating the bench it is focusing
it’s generating the bench it is focusing
it’s generating the bench it is focusing its attention on bench or a zebra
its attention on bench or a zebra
its attention on bench or a zebra standing next to a zebra in a dirt field
standing next to a zebra in a dirt field
standing next to a zebra in a dirt field these are very factual descriptions
these are very factual descriptions
these are very factual descriptions we’re not gonna get very interesting
we’re not gonna get very interesting
we’re not gonna get very interesting ones from them but indeed the first
ones from them but indeed the first
ones from them but indeed the first zebra it’s focusing its attention on is
zebra it’s focusing its attention on is
zebra it’s focusing its attention on is the one in the foreground and then to
the one in the foreground and then to
the one in the foreground and then to the next zebra it’s actually the one in
the next zebra it’s actually the one in
the next zebra it’s actually the one in the background and you see that color
the background and you see that color
the background and you see that color coded too so computer vision algorithms
coded too so computer vision algorithms
coded too so computer vision algorithms have gotten a lot more sophisticated and
have gotten a lot more sophisticated and
have gotten a lot more sophisticated and actually also telling us a little bit
actually also telling us a little bit
actually also telling us a little bit where they’re paying attention to as
where they’re paying attention to as
where they’re paying attention to as they’re trying to translate from visual
they’re trying to translate from visual
they’re trying to translate from visual data into text it gets even further we
data into text it gets even further we
data into text it gets even further we can also do so-called visual question
can also do so-called visual question
can also do so-called visual question answering this is an interesting task
answering this is an interesting task
answering this is an interesting task where you basically as training data
where you basically as training data
where you basically as training data give it an image a question and an
give it an image a question and an
give it an image a question and an answer and now you want to probe the
answer and now you want to probe the
answer and now you want to probe the algorithm you can ask it lots of the
algorithm you can ask it lots of the
algorithm you can ask it lots of the some kinds of questions about an image
some kinds of questions about an image
some kinds of questions about an image and you see if it still gets dried or
and you see if it still gets dried or
and you see if it still gets dried or not so the example here at the top what
not so the example here at the top what
not so the example here at the top what color are the bananas is a good one
color are the bananas is a good one
color are the bananas is a good one because if you didn’t know the image in
because if you didn’t know the image in
because if you didn’t know the image in 90% of the cases you’d probably be
90% of the cases you’d probably be
90% of the cases you’d probably be correct just saying yellow without the
correct just saying yellow without the
correct just saying yellow without the image but we have here the visualization
image but we have here the visualization
image but we have here the visualization also of where the algorithm is paying
also of where the algorithm is paying
also of where the algorithm is paying attention to as it’s trying to answer
attention to as it’s trying to answer
attention to as it’s trying to answer this particular question and it’s
this particular question and it’s
this particular question and it’s actually focusing its attention on the
actually focusing its attention on the
actually focusing its attention on the brighter areas in that image and
brighter areas in that image and
brighter areas in that image and realizes those bananas are actually
realizes those bananas are actually
realizes those bananas are actually green and gives the correct answer
green and gives the correct answer
green and gives the correct answer another fun one what is the pattern on
another fun one what is the pattern on
another fun one what is the pattern on the cat’s fur on its tail it actually
the cat’s fur on its tail it actually
the cat’s fur on its tail it actually focuses most of its attention again in
focuses most of its attention again in
focuses most of its attention again in that bright area on the cat’s tail and
that bright area on the cat’s tail and
that bright area on the cat’s tail and correctly identifies this as striped and
correctly identifies this as striped and
correctly identifies this as striped and another fun example what is the boy
another fun example what is the boy
another fun example what is the boy holding it’s actually figuring out based
holding it’s actually figuring out based
holding it’s actually figuring out based on the question that you now need to
on the question that you now need to
on the question that you now need to focus your attention on the arm and the
focus your attention on the arm and the
focus your attention on the arm and the object below the arm and correctly
object below the arm and correctly
object below the arm and correctly classifies this as surfboard so those
classifies this as surfboard so those
classifies this as surfboard so those are some of the great applications that
are some of the great applications that
are some of the great applications that we’re now able to do as long as we have
we’re now able to do as long as we have
we’re now able to do as long as we have enough training data about a certain
enough training data about a certain
enough training data about a certain domain if you never show at a baseball
domain if you never show at a baseball
domain if you never show at a baseball image and it’s training time it won’t be
image and it’s training time it won’t be
image and it’s training time it won’t be able to answer any questions about
able to answer any questions about
able to answer any questions about baseball they actually even more
baseball they actually even more
baseball they actually even more powerful applications that we can now do
powerful applications that we can now do
powerful applications that we can now do with computer vision one of the ones
with computer vision one of the ones
with computer vision one of the ones that I’m very excited about as in the
that I’m very excited about as in the
that I’m very excited about as in the medical field particular oncology this
medical field particular oncology this
medical field particular oncology this is a small start-up a felis that
is a small start-up a felis that
is a small start-up a felis that actually is automating blood cell
actually is automating blood cell
actually is automating blood cell counting so you can make it very small
counting so you can make it very small
counting so you can make it very small pricking your finger and you can count
pricking your finger and you can count
pricking your finger and you can count blood cells with the same kind of
blood cells with the same kind of
blood cells with the same kind of architecture as the one I showed you
architecture as the one I showed you
architecture as the one I showed you before it’s a convolutional neural
before it’s a convolutional neural
before it’s a convolutional neural network and now that you make this so
network and now that you make this so
network and now that you make this so cheap this used to cost a couple hundred
cheap this used to cost a couple hundred
cheap this used to cost a couple hundred dollars to have people actually sit
dollars to have people actually sit
dollars to have people actually sit there and for each blood sample count
there and for each blood sample count
there and for each blood sample count how many red or white blood cells they
how many red or white blood cells they
how many red or white blood cells they are now that you can make this much
are now that you can make this much
are now that you can make this much cheaper you can prove oncology care you
cheaper you can prove oncology care you
cheaper you can prove oncology care you can identify infections and help
can identify infections and help
can identify infections and help patients with leukemia and so on in
patients with leukemia and so on in
patients with leukemia and so on in general I think radiology will also have
general I think radiology will also have
general I think radiology will also have a huge impact with AI the problem with
a huge impact with AI the problem with
a huge impact with AI the problem with radiology is that we need a lot of
radiology is that we need a lot of
radiology is that we need a lot of trained data because unlike in a blood
trained data because unlike in a blood
trained data because unlike in a blood scan or pathology scan you’re looking
scan or pathology scan you’re looking
scan or pathology scan you’re looking for a thousand different things that
for a thousand different things that
for a thousand different things that could be wrong in a head CT
could be wrong in a head CT
could be wrong in a head CT a scan and it will take us a while
a scan and it will take us a while
a scan and it will take us a while before we could automate that entire
before we could automate that entire
before we could automate that entire process so for a very long time
process so for a very long time
process so for a very long time AI will work together with radiologists
AI will work together with radiologists
AI will work together with radiologists to improve that process and in fact we
to improve that process and in fact we
to improve that process and in fact we already know that we can identify
already know that we can identify
already know that we can identify certain things that can very quickly
certain things that can very quickly
certain things that can very quickly kill you so for instance a stroke or a
kill you so for instance a stroke or a
kill you so for instance a stroke or a so called intracranial hemorrhage brain
so called intracranial hemorrhage brain
so called intracranial hemorrhage brain bleeds those we can identify very
bleeds those we can identify very
bleeds those we can identify very quickly and then without knowing all the
quickly and then without knowing all the
quickly and then without knowing all the other things that might be wrong in a
other things that might be wrong in a
other things that might be wrong in a head CT scan we can put those to the top
head CT scan we can put those to the top
head CT scan we can put those to the top of the queue in an emergency room
of the queue in an emergency room
of the queue in an emergency room setting and that can already save lives
setting and that can already save lives
setting and that can already save lives now we talked about computer vision and
now we talked about computer vision and
now we talked about computer vision and speech recognition as two successes of
speech recognition as two successes of
speech recognition as two successes of AI there’s actually still some areas
AI there’s actually still some areas
AI there’s actually still some areas that we’re struggling with and that is
that we’re struggling with and that is
that we’re struggling with and that is motor control this is a DARPA Grand
motor control this is a DARPA Grand
motor control this is a DARPA Grand Challenge and a bunch of examples of
Challenge and a bunch of examples of
Challenge and a bunch of examples of some very expensive robots that are
some very expensive robots that are
some very expensive robots that are trying to walk around open doors turn
trying to walk around open doors turn
trying to walk around open doors turn levers and things like that and as you
levers and things like that and as you
levers and things like that and as you can see we’re still quite far away as a
can see we’re still quite far away as a
can see we’re still quite far away as a community in fact you could say that
community in fact you could say that
community in fact you could say that we’re not even at the level of Abby yet
we’re not even at the level of Abby yet
we’re not even at the level of Abby yet when it comes to motor control abby is
when it comes to motor control abby is
when it comes to motor control abby is actually quite complex has a million
actually quite complex has a million
actually quite complex has a million neurons it needs to identify a lot of
neurons it needs to identify a lot of
neurons it needs to identify a lot of different paths and so on and we’re not
different paths and so on and we’re not
different paths and so on and we’re not there yet so that is a very active area
there yet so that is a very active area
there yet so that is a very active area of research one of the most interesting
of research one of the most interesting
of research one of the most interesting manifestations of human intelligence I
manifestations of human intelligence I
manifestations of human intelligence I think is language and in language making
think is language and in language making
think is language and in language making a huge amount of progress right now but
a huge amount of progress right now but
a huge amount of progress right now but there’s still a lot of ways to go so
there’s still a lot of ways to go so
there’s still a lot of ways to go so here’s an example I think we could bed
here’s an example I think we could bed
here’s an example I think we could bed do better now but this is from 2011 when
do better now but this is from 2011 when
do better now but this is from 2011 when folks realized that whenever Anne
folks realized that whenever Anne
folks realized that whenever Anne Hathaway famous actress won a couple of
Hathaway famous actress won a couple of
Hathaway famous actress won a couple of Oscars starred in a movie the reviews
Oscars starred in a movie the reviews
Oscars starred in a movie the reviews came out all the sudden the stocks for
came out all the sudden the stocks for
came out all the sudden the stocks for the company Berkshire Hathaway go up a
the company Berkshire Hathaway go up a
the company Berkshire Hathaway go up a significant amount so it was already
significant amount so it was already
significant amount so it was already clear in 2011 people were trying to use
clear in 2011 people were trying to use
clear in 2011 people were trying to use natural language processing for
natural language processing for
natural language processing for algorithmic trading and in this case
algorithmic trading and in this case
algorithmic trading and in this case made the mistake of so-called entity
made the mistake of so-called entity
made the mistake of so-called entity disambiguation they disambiguated a
disambiguation they disambiguated a
disambiguation they disambiguated a Hathaway to the company instead of the
Hathaway to the company instead of the
Hathaway to the company instead of the actress and then made pretty substantial
actress and then made pretty substantial
actress and then made pretty substantial monetary decisions
where we have gotten better an NLP is
where we have gotten better an NLP is actually on just text classification in
actually on just text classification in
actually on just text classification in fact here are a couple of examples of
fact here are a couple of examples of
fact here are a couple of examples of sentences that up until two years ago
sentences that up until two years ago
sentences that up until two years ago pretty much every algorithm out there
pretty much every algorithm out there
pretty much every algorithm out there would have incorrectly classified the
would have incorrectly classified the
would have incorrectly classified the first sentence is in its ragged cheap
first sentence is in its ragged cheap
first sentence is in its ragged cheap and unassuming way the movie works so
and unassuming way the movie works so
and unassuming way the movie works so traditional algorithms would have said
traditional algorithms would have said
traditional algorithms would have said well it’s ragged and cheap so it’s
well it’s ragged and cheap so it’s
well it’s ragged and cheap so it’s probably negative sentence because they
probably negative sentence because they
probably negative sentence because they haven’t had the ability to capture the
haven’t had the ability to capture the
haven’t had the ability to capture the whole context but now and what you see
whole context but now and what you see
whole context but now and what you see here is we actually have two passes over
here is we actually have two passes over
here is we actually have two passes over that sentence and in the second pass the
that sentence and in the second pass the
that sentence and in the second pass the algorithm focuses much more on works the
algorithm focuses much more on works the
algorithm focuses much more on works the movie actually working despite its flaws
movie actually working despite its flaws
movie actually working despite its flaws so it correctly classified this as
so it correctly classified this as
so it correctly classified this as positive and the second one is the
positive and the second one is the
positive and the second one is the opposite kind of example the best way to
opposite kind of example the best way to
opposite kind of example the best way to hope for any chance of enjoying this
hope for any chance of enjoying this
hope for any chance of enjoying this film is by lowering your expectations
film is by lowering your expectations
film is by lowering your expectations again a kind of example that is quite
again a kind of example that is quite
again a kind of example that is quite tricky because algorithms in the past
tricky because algorithms in the past
tricky because algorithms in the past would just look at single words and has
would just look at single words and has
would just look at single words and has best and hope and chance and enjoying
best and hope and chance and enjoying
best and hope and chance and enjoying such a positive sentence yet you can
such a positive sentence yet you can
such a positive sentence yet you can only get there if you already think the
only get there if you already think the
only get there if you already think the movie is pretty crappy so those are
movie is pretty crappy so those are
movie is pretty crappy so those are examples that we can now do largely
examples that we can now do largely
examples that we can now do largely because of advances also again in deep
because of advances also again in deep
because of advances also again in deep learning there’s some active areas of
learning there’s some active areas of
learning there’s some active areas of research that we still work on and one
research that we still work on and one
research that we still work on and one of them is text summarization it’s
of them is text summarization it’s
of them is text summarization it’s actually a really tricky problem pretty
actually a really tricky problem pretty
actually a really tricky problem pretty much every natural language processing
much every natural language processing
much every natural language processing in a model that you’ve seen in the past
in a model that you’ve seen in the past
in a model that you’ve seen in the past only can generate at most a sentence
only can generate at most a sentence
only can generate at most a sentence coherently once tried when we tried as a
coherently once tried when we tried as a
coherently once tried when we tried as a community to generate longer sequences
community to generate longer sequences
community to generate longer sequences fully automatically in this end-to-end
fully automatically in this end-to-end
fully automatically in this end-to-end deep learning models which usually
deep learning models which usually
deep learning models which usually didn’t do very well so this is a result
didn’t do very well so this is a result
didn’t do very well so this is a result from just a couple of months ago where
from just a couple of months ago where
from just a couple of months ago where our group worked on summarization and
our group worked on summarization and
our group worked on summarization and you see here at the bottom a longer
you see here at the bottom a longer
you see here at the bottom a longer document and at the top you see the
document and at the top you see the
document and at the top you see the summary and what’s fascinating here is
summary and what’s fascinating here is
summary and what’s fascinating here is that it actually the summarization
that it actually the summarization
that it actually the summarization algorithm learned to some cases copy and
algorithm learned to some cases copy and
algorithm learned to some cases copy and paste particular words sometimes entire
paste particular words sometimes entire
paste particular words sometimes entire phrases but sometimes it also picks and
phrases but sometimes it also picks and
phrases but sometimes it also picks and chooses which words to pick from which
chooses which words to pick from which
chooses which words to pick from which area of the longer document in order to
area of the longer document in order to
area of the longer document in order to generate the summary in many cases
generate the summary in many cases
generate the summary in many cases actually generates coherent longer
actually generates coherent longer
actually generates coherent longer document summarize summaries and as the
document summarize summaries and as the
document summarize summaries and as the summary correctly says to do this really
summary correctly says to do this really
summary correctly says to do this really well still remains an open research
well still remains an open research
well still remains an open research problem one of the areas of NLP that I’m
problem one of the areas of NLP that I’m
problem one of the areas of NLP that I’m personally most excited about is
personally most excited about is
personally most excited about is question answering because you can
question answering because you can
question answering because you can actually think of question answering as
actually think of question answering as
actually think of question answering as a task that encompasses pretty much
a task that encompasses pretty much
a task that encompasses pretty much every other NLP task you can ask what is
every other NLP task you can ask what is
every other NLP task you can ask what is the translation of the sentence into
the translation of the sentence into
the translation of the sentence into French you can ask what is the sentiment
French you can ask what is the sentiment
French you can ask what is the sentiment you can ask what is the summary right in
you can ask what is the summary right in
you can ask what is the summary right in some ways everything becomes a question
some ways everything becomes a question
some ways everything becomes a question answering problem if you have a really
answering problem if you have a really
answering problem if you have a really powerful question answering model and so
powerful question answering model and so
powerful question answering model and so here we worked on this regrade data set
here we worked on this regrade data set
here we worked on this regrade data set that stanford collected called squad
that stanford collected called squad
that stanford collected called squad with stanford question answering data
with stanford question answering data
with stanford question answering data set that takes lots of wikipedia
set that takes lots of wikipedia
set that takes lots of wikipedia articles and then asks crowd workers to
articles and then asks crowd workers to
articles and then asks crowd workers to collect questions and then also
collect questions and then also
collect questions and then also different guard workers to collect the
different guard workers to collect the
different guard workers to collect the answers and the models that you now see
answers and the models that you now see
answers and the models that you now see that dwell on this task are much more
that dwell on this task are much more
that dwell on this task are much more complex than a model I showed you in the
complex than a model I showed you in the
complex than a model I showed you in the beginning that had the same kind of
beginning that had the same kind of
beginning that had the same kind of layer just learning more abstract
layer just learning more abstract
layer just learning more abstract representations language by itself also
representations language by itself also
representations language by itself also seems to require a lot of distributed
seems to require a lot of distributed
seems to require a lot of distributed computing in our brain that takes a lot
computing in our brain that takes a lot
computing in our brain that takes a lot of different parts and elements now what
of different parts and elements now what
of different parts and elements now what makes me really excited and looking
makes me really excited and looking
makes me really excited and looking forward to this to the next couple of
forward to this to the next couple of
forward to this to the next couple of years is the number of people that are
years is the number of people that are
years is the number of people that are now entering the field it’s there’s a
now entering the field it’s there’s a
now entering the field it’s there’s a lot of excitement in AI and just in a
lot of excitement in AI and just in a
lot of excitement in AI and just in a class I echo talked with Chris Manning
class I echo talked with Chris Manning
class I echo talked with Chris Manning earlier this year we had over 660
earlier this year we had over 660
earlier this year we had over 660 students at Stanford attending that
students at Stanford attending that
students at Stanford attending that class even though it’s a graduate level
class even though it’s a graduate level
class even though it’s a graduate level class and there are hundreds of
class and there are hundreds of
class and there are hundreds of thousands of views on YouTube of pretty
thousands of views on YouTube of pretty
thousands of views on YouTube of pretty technical material and that is very
technical material and that is very
technical material and that is very exciting but as we see AI actually
exciting but as we see AI actually
exciting but as we see AI actually working we have to acknowledge also that
working we have to acknowledge also that
working we have to acknowledge also that it’s just a tool and it will stay a tool
it’s just a tool and it will stay a tool
it’s just a tool and it will stay a tool for the foreseeable future we don’t
for the foreseeable future we don’t
for the foreseeable future we don’t really have to worry about Skynet or
really have to worry about Skynet or
really have to worry about Skynet or terminator kinds of scenarios but what
terminator kinds of scenarios but what
terminator kinds of scenarios but what is important is to understand that tools
is important is to understand that tools
is important is to understand that tools can be used in good ways and in bad ways
can be used in good ways and in bad ways
can be used in good ways and in bad ways in some ways AI is just like the
in some ways AI is just like the
in some ways AI is just like the internet or hammer or cars and you can
internet or hammer or cars and you can
internet or hammer or cars and you can use them as weapons or you can use them
use them as weapons or you can use them
use them as weapons or you can use them to transport sick people and it’s
to transport sick people and it’s
to transport sick people and it’s important for us to acknowledge that
important for us to acknowledge that
important for us to acknowledge that the tools are only as good as the people
the tools are only as good as the people
the tools are only as good as the people and the political systems that end up
and the political systems that end up
and the political systems that end up using them in fact if we use them well I
using them in fact if we use them well I
using them in fact if we use them well I think AI and especially AI power
think AI and especially AI power
think AI and especially AI power language capabilities will allow us to
language capabilities will allow us to
language capabilities will allow us to improve our communication pretty sure we
improve our communication pretty sure we
improve our communication pretty sure we can eventually have the Babel Fish in
can eventually have the Babel Fish in
can eventually have the Babel Fish in the next couple of years where we talk
the next couple of years where we talk
the next couple of years where we talk in one language and we listen in another
in one language and we listen in another
in one language and we listen in another one coming live at the other end we can
one coming live at the other end we can
one coming live at the other end we can improve access to information question
improve access to information question
improve access to information question answering is a great example for that
answering is a great example for that
answering is a great example for that and in general make work much more
and in general make work much more
and in general make work much more efficient in fact I think we’ll be able
efficient in fact I think we’ll be able
efficient in fact I think we’ll be able to automate most of the basic human
to automate most of the basic human
to automate most of the basic human needs like food we can automate farming
needs like food we can automate farming
needs like food we can automate farming with computer vision and some simpler
with computer vision and some simpler
with computer vision and some simpler robotic control we can build houses
robotic control we can build houses
robotic control we can build houses automatically and so on I think in the
automatically and so on I think in the
automatically and so on I think in the end as human intelligence and
end as human intelligence and
end as human intelligence and productivity gets enhanced I hope that
productivity gets enhanced I hope that
productivity gets enhanced I hope that that will lead us to a future where we
that will lead us to a future where we
that will lead us to a future where we can focus on unique and creative tasks
can focus on unique and creative tasks
can focus on unique and creative tasks and those kinds of tasks that require
and those kinds of tasks that require
and those kinds of tasks that require empathy and where we care for each other
empathy and where we care for each other
empathy and where we care for each other and we can basically automate a lot of
and we can basically automate a lot of
and we can basically automate a lot of the boring treachery that is out there
the boring treachery that is out there
the boring treachery that is out there what’s important to acknowledge is that
what’s important to acknowledge is that
what’s important to acknowledge is that AI is only as good as a training data
AI is only as good as a training data
AI is only as good as a training data that we give it if your training data is
that we give it if your training data is
that we give it if your training data is sexist or racist then the I will pick
sexist or racist then the I will pick
sexist or racist then the I will pick those patterns up and in some cases
those patterns up and in some cases
those patterns up and in some cases repeat them or even amplify them so as
repeat them or even amplify them so as
repeat them or even amplify them so as we’re applying AI to more and more
we’re applying AI to more and more
we’re applying AI to more and more different areas in simple things like
different areas in simple things like
different areas in simple things like loan applications but also more complex
loan applications but also more complex
loan applications but also more complex things like the financial system or the
things like the financial system or the
things like the financial system or the judicial system and medical applications
judicial system and medical applications
judicial system and medical applications and political advertisement and so on
and political advertisement and so on
and political advertisement and so on they I will eventually be in all of
they I will eventually be in all of
they I will eventually be in all of these areas I think it’s important that
these areas I think it’s important that
these areas I think it’s important that we think about regulations or at least
we think about regulations or at least
we think about regulations or at least guidelines to prevent the negative
guidelines to prevent the negative
guidelines to prevent the negative effects that could happen and that may
effects that could happen and that may
effects that could happen and that may have already been in our training data
have already been in our training data
have already been in our training data and lastly it’s not just the data that
and lastly it’s not just the data that
and lastly it’s not just the data that might have biases is also the communion
might have biases is also the communion
might have biases is also the communion itself and the IEEE community right now
itself and the IEEE community right now
itself and the IEEE community right now it’s actually quite aware that we have a
it’s actually quite aware that we have a
it’s actually quite aware that we have a diversity issue and that is some
diversity issue and that is some
diversity issue and that is some that we continue to work on all right
that we continue to work on all right
that we continue to work on all right thank you
thank you
thank you [Applause]
”
Be First to Comment