and feeling that there is quite some
confusion when we talk about artificial
intelligence I mean for instance look at
this picture already
is this why did they choose a jellyfish
to represent intelligence I was relieved
to know that this was a particularly
intelligent jellyfish but in general
artificial intelligence is creating all
sort of fears in people so current AI
colour and artificial intelligence is
creating all sorts of conversations all
sort of dialogue and dialogue between
people from self-driving cars to for
instance the future of humanity
coexisting with intelligent machines so
what do you really feel about it you
feel scared do you feel worried do you
feel enthusiastic do you feel curious in
other terms what comes to your mind when
you hear the term artificial
intelligence one thing is actually easy
to say the term artificial intelligence
is just a bad choice to describe what is
happening today and I’m sure that
Stanley Kubrick would agree with me if I
tell you to just forget about Space
Odyssey or any other science fiction
science fiction movie you might have
seen recently about AI about the tips on
intelligence actually I should say that
a lot of the predictions contained in
this movie Space Oddity
Odyssey became true for instance the
iPad or the Alavi L automatic
translation but not the part concerning
intelligence the word artificial
intelligence goes back to the 50s to the
very beginning of computer science and
essentially no progress has been made
since then so current AI is not about
creating thinking machine is about is
about inventing our cognitive capability
quite a different matter
so today when we speak we talk about
artificial intelligence what we really
mean is machine learning
and also machine learning is not new
most of the results go back to 30 or 40
years ago however what is new is that in
the last decade machine learning has
experienced a spectacular increase in
performance thanks to the combined
effect of new computational tools huge
and rich datasets and new algorithms
it’s quite important to observe that
once machine learning reaches certain
capabilities reaches the capability of
having human accuracy in identifying an
object in a picture or any other
computational task then it can do it at
a massive scale and in this sense become
super human so machine learning is
certainly not just one single techniques
there are many different techniques that
are used in many different context
however most of the recent progress has
been triggered by one such tool that we
this called deep learning and I will
discuss this in a minute but before
doing that I would like to clarify one
point machine learning is about the
capability of extracting information
from data and if you think about it this
is a very important task is a kind of a
precondition for intelligence even
before starting to think about creating
a human-like thinking machine you need
to be able to create a machine that is
capable of giving a meaning to the data
it receives to the input it receives
from the external world and that
continuously learns from them
this is totally natural for animals
thanks to the evolution of their brain
and all of their visual system they can
instantaneously detect any anomaly any
the presence of any object in a visual
scene but this has always been very very
difficult for a computer let’s try it
before we enter into discussions of how
machine learning works let’s try to
write the computer code that you know
performs object recognition for instance
suppose you want to write the code that
recognises the presence of a chair in an
image then what you would need to do is
to write few lines suppose that you are
a software engineer with few line of
codes which say something like if the
object in the picture has four legs or
if it has a flat surface and if it has a
flat surface to sit on or if it has a
backrest then this is a chair and this
would perfectly work for the first
example of chair in this slide
all right however already for the second
chair is not going to work and the
reason is that it doesn’t have four legs
so you would have to go back and modify
your code and write a line which says
well if it has four legs or five wheels
then this is going to be a chair for the
other conditions are met but again if
you show a chair produced by a designer
then you get completely lost you don’t
even know how to encode this so we need
a completely different strategy we need
to learn from examples rather than to
the final priori what the objects are
and this is what machine learning is
about now once a machine learns to
recognize an object then at a human
human accuracy then it becomes super
human in which sense in the sense that
can do it at a massive scale so this let
me discuss an example of an application
suppose you want to do breast cancer
screening in your region and what you
would have to do is that each year you
would have to analyze and detect
anomalies in millions and millions of
mammographic images and you just don’t
have enough radiologist for doing that
and I doubt that radiologists would
really like to do that as a full-time
job so what you really need is like a
machine that with human accuracy is
capable of detecting the presence of
anomalies and this is actually doable
today so the machines that are capable
of doing this
a machine that learn from previous human
information namely thousands of images
that have been studied annotated
classified by human radiology
and then building on this information
they become superhuman at analyzing new
images so as I said machine learning is
a set of tools and I cannot describe all
of them I’m just going to focus on on
the single tool that has triggered most
of the interest in in this last year and
this is called this technique is called
deep learning these are neural networks
that were perform a particular type of
learning and I will try to discuss this
a deep neural network is composed of
many layers you have an input layer and
an output layer and then you have many
intermediate layers that process the
information so the information enters on
the left and goes out on the right for
instance if you plug in the image of a
chair you want to read out in the output
layer an answer which says there is a
chair so the output layer is going to
tell you how the network classifies the
input so each network is composed of
many layers each layer is composed of
thousands of nodes that are called a
little artificial neurons they behave as
stylized model of biological neurons and
then the overall network is composed of
millions and millions of connections the
connection between neurons can be either
weak or strong depending on how much to
connect neurons influence each other now
this picture is just like a pinhead in
this room so real deep network can have
you know tens of millions of connections
so there are huge huge objects so how
does learning take place in this kind of
systems well learning take place in in a
rather simple manner you provide a lot
of examples to this network of inputs
and the corresponding output and then
you progressively and I terribly change
all the connection strength up to the
point in which the network is capable of
classifying correctly all the examples
that
provided so you have an eidetic process
in which you change in an automatic
manner millions of connection presenting
million of examples and you repeat this
operation million of times so it’s a
quite heavy process and this is why all
these research activity has been
triggered by the possibility of having a
lot of computational power now once the
network has learned to classify all the
examples for instance angles all phases
then it becomes help then you can stop
and use it to make predictions and that
point becomes useful it has learned to
classify the all the examples and the
point you stop and you feed a new image
for instance that the network has never
seen before and you’re going to read the
answer and this is going to be a
prediction and this machine that have
demonstrated to be particularly
efficient in doing this kind of task
they actually have superhuman
performance in the sense that not only
they can do it at a massive scale but
even the accuracy with which they do it
is very high and along the learning
process the values layer specialized in
doing different things the first layers
become specialized in extracting all the
fundamental features that characterize
the input and the last layer perform the
classification for instance this image
is capable of recognizes a picture of me
25 years ago without up never having
never seen this picture ok and this
would and would be able to do it for
essentially all of you now there’s
nothing special about images all these
techniques can be applied to any field
of science human activities in which
data are present and are important
therefore this is going to affect our
future in a 360 degrees all our activity
which we need to process information are
going to change we know that jobs are
going to change and society is going to
change so it’s very important that we
understand what we are talking about in
this context in the case of basic
science for instance you could do
nowadays these techniques are
used for analyzing data that lead to
climate changes they are neat used to
optimize the process of gene editing a
very advanced molecular biology
technique they used to reconstruct the
three-dimensional structure of the of
the brain stantly we still don’t know
and it’s hopefully going to to be to
have a huge impact and it’s even used in
astrophysics so the whole spectrum of of
science is is included and I’m not
mentioning you know all the application
in in human in daily human life this is
all fine however as performant as all
these things may appear I should say
that we are only at a very preliminary
stage there are open problems and
challenges ahead of us for instance
about the functioning of these kinds of
devices we still need to understand how
they how learning can be made faster how
the prediction performance can be
improved and act at a more fundamental
level we need to understand what makes
this kind of devices special compared to
other type of devices so where does the
propensity to learn and generalize come
from these are mathematical question
that I don’t want to trace here but
these are open problems which we all
work then our challenges let me mention
two challenges that are very important
and they’re going to give you a clear
idea of what could be the impact in the
future of this kind of technology the
first challenge is to learn to learn
without supervision so more the vast
majority of the data in the world cannot
be annotated classified by humans and
they contain a lot of information that
cannot even be understood by humans okay
so we need to develop algorithm that are
algorithms that are capable to extract
information to identify patterns to make
prediction without human supervision ok
so this is you know a challenge and all
the research which I have just started
to work on these are working on this and
it’s totally open challenge
another critical objective is these kind
of networks can
predict cannot identify the
relationships between causes and effects
in data so somehow this machine have to
become capable of identifying causation
patterns in in the data in the raw data
when this will become possible then this
tool will provide a lot of information
about modeling systems and making very
complicated prediction and making you
know new tools available as an example
of this I mean you can use the tip
network to recognize images – for speech
recognition you can use it for playing
games and so on and you can use it for
automatic translation you can translate
from a language to another language that
are similar and this translation is just
based on statistics but you cannot use
this kind of tools to create any new
type of sentence because they don’t get
the logical relationship between words
okay just try to set up a conversation
between Siri and Google assistant and
you will see that the conversation
becomes totally banal and boring
immediately as a researcher however all
these kind of open problems are actually
exciting we don’t get depressed when we
find open problems we actually get
excited and and I think they give the
reason is that they clearly show that
there is a huge space for improvement
and for and for application across all
disciplines so there are some
fundamental challenges that have you
know there’s a huge space for
improvement a lot of space for research
and a lot of consequences ahead of us so
what should a student study in order to
be prepared to contribute to this field
this is a question that sometimes occurs
to my mind and actually quite often and
what the answer in my opinion is simple
and I have to say that a lot of
colleagues of mine agree
these there’s no shortcut students need
to know the fundamentals they need to
know math they need to know physics they
need to know computer science they need
to know biology they need to know
chemistry they need to know economics
they didn’t to have all the
methodological tools that allow them to
analyze complex systems they need to
have this background this is the only
way they can actually use the cognitive
power provided by these new tools so the
message for students is actually do not
specialize too soon you have to become
strong from a methodological point of
view be able to model complex system
understand all the tools and then use
this tool or even advance this tool so I
hope that this discussion is somehow
clarified which is the current state of
the art in artificial intelligence the
present of artificial intelligence is
not about thinking machine machines is
about augmenting our cognitive
capabilities through this tool that are
already at work and that have an immense
possibility of making further progress
so as a researcher I would really like
to sit on the other side of the stage
together with the students and the
reason is that there are fantastic and
broad challenges ahead of us and
therefore this is a truly remarkable
time to be a student [Applause]
confusion when we talk about artificial
intelligence I mean for instance look at
this picture already
is this why did they choose a jellyfish
to represent intelligence I was relieved
to know that this was a particularly
intelligent jellyfish but in general
artificial intelligence is creating all
sort of fears in people so current AI
colour and artificial intelligence is
creating all sorts of conversations all
sort of dialogue and dialogue between
people from self-driving cars to for
instance the future of humanity
coexisting with intelligent machines so
what do you really feel about it you
feel scared do you feel worried do you
feel enthusiastic do you feel curious in
other terms what comes to your mind when
you hear the term artificial
intelligence one thing is actually easy
to say the term artificial intelligence
is just a bad choice to describe what is
happening today and I’m sure that
Stanley Kubrick would agree with me if I
tell you to just forget about Space
Odyssey or any other science fiction
science fiction movie you might have
seen recently about AI about the tips on
intelligence actually I should say that
a lot of the predictions contained in
this movie Space Oddity
Odyssey became true for instance the
iPad or the Alavi L automatic
translation but not the part concerning
intelligence the word artificial
intelligence goes back to the 50s to the
very beginning of computer science and
essentially no progress has been made
since then so current AI is not about
creating thinking machine is about is
about inventing our cognitive capability
quite a different matter
so today when we speak we talk about
artificial intelligence what we really
mean is machine learning
and also machine learning is not new
most of the results go back to 30 or 40
years ago however what is new is that in
the last decade machine learning has
experienced a spectacular increase in
performance thanks to the combined
effect of new computational tools huge
and rich datasets and new algorithms
it’s quite important to observe that
once machine learning reaches certain
capabilities reaches the capability of
having human accuracy in identifying an
object in a picture or any other
computational task then it can do it at
a massive scale and in this sense become
super human so machine learning is
certainly not just one single techniques
there are many different techniques that
are used in many different context
however most of the recent progress has
been triggered by one such tool that we
this called deep learning and I will
discuss this in a minute but before
doing that I would like to clarify one
point machine learning is about the
capability of extracting information
from data and if you think about it this
is a very important task is a kind of a
precondition for intelligence even
before starting to think about creating
a human-like thinking machine you need
to be able to create a machine that is
capable of giving a meaning to the data
it receives to the input it receives
from the external world and that
continuously learns from them
this is totally natural for animals
thanks to the evolution of their brain
and all of their visual system they can
instantaneously detect any anomaly any
the presence of any object in a visual
scene but this has always been very very
difficult for a computer let’s try it
before we enter into discussions of how
machine learning works let’s try to
write the computer code that you know
performs object recognition for instance
suppose you want to write the code that
recognises the presence of a chair in an
image then what you would need to do is
to write few lines suppose that you are
a software engineer with few line of
codes which say something like if the
object in the picture has four legs or
if it has a flat surface and if it has a
flat surface to sit on or if it has a
backrest then this is a chair and this
would perfectly work for the first
example of chair in this slide
all right however already for the second
chair is not going to work and the
reason is that it doesn’t have four legs
so you would have to go back and modify
your code and write a line which says
well if it has four legs or five wheels
then this is going to be a chair for the
other conditions are met but again if
you show a chair produced by a designer
then you get completely lost you don’t
even know how to encode this so we need
a completely different strategy we need
to learn from examples rather than to
the final priori what the objects are
and this is what machine learning is
about now once a machine learns to
recognize an object then at a human
human accuracy then it becomes super
human in which sense in the sense that
can do it at a massive scale so this let
me discuss an example of an application
suppose you want to do breast cancer
screening in your region and what you
would have to do is that each year you
would have to analyze and detect
anomalies in millions and millions of
mammographic images and you just don’t
have enough radiologist for doing that
and I doubt that radiologists would
really like to do that as a full-time
job so what you really need is like a
machine that with human accuracy is
capable of detecting the presence of
anomalies and this is actually doable
today so the machines that are capable
of doing this
a machine that learn from previous human
information namely thousands of images
that have been studied annotated
classified by human radiology
and then building on this information
they become superhuman at analyzing new
images so as I said machine learning is
a set of tools and I cannot describe all
of them I’m just going to focus on on
the single tool that has triggered most
of the interest in in this last year and
this is called this technique is called
deep learning these are neural networks
that were perform a particular type of
learning and I will try to discuss this
a deep neural network is composed of
many layers you have an input layer and
an output layer and then you have many
intermediate layers that process the
information so the information enters on
the left and goes out on the right for
instance if you plug in the image of a
chair you want to read out in the output
layer an answer which says there is a
chair so the output layer is going to
tell you how the network classifies the
input so each network is composed of
many layers each layer is composed of
thousands of nodes that are called a
little artificial neurons they behave as
stylized model of biological neurons and
then the overall network is composed of
millions and millions of connections the
connection between neurons can be either
weak or strong depending on how much to
connect neurons influence each other now
this picture is just like a pinhead in
this room so real deep network can have
you know tens of millions of connections
so there are huge huge objects so how
does learning take place in this kind of
systems well learning take place in in a
rather simple manner you provide a lot
of examples to this network of inputs
and the corresponding output and then
you progressively and I terribly change
all the connection strength up to the
point in which the network is capable of
classifying correctly all the examples
that
provided so you have an eidetic process
in which you change in an automatic
manner millions of connection presenting
million of examples and you repeat this
operation million of times so it’s a
quite heavy process and this is why all
these research activity has been
triggered by the possibility of having a
lot of computational power now once the
network has learned to classify all the
examples for instance angles all phases
then it becomes help then you can stop
and use it to make predictions and that
point becomes useful it has learned to
classify the all the examples and the
point you stop and you feed a new image
for instance that the network has never
seen before and you’re going to read the
answer and this is going to be a
prediction and this machine that have
demonstrated to be particularly
efficient in doing this kind of task
they actually have superhuman
performance in the sense that not only
they can do it at a massive scale but
even the accuracy with which they do it
is very high and along the learning
process the values layer specialized in
doing different things the first layers
become specialized in extracting all the
fundamental features that characterize
the input and the last layer perform the
classification for instance this image
is capable of recognizes a picture of me
25 years ago without up never having
never seen this picture ok and this
would and would be able to do it for
essentially all of you now there’s
nothing special about images all these
techniques can be applied to any field
of science human activities in which
data are present and are important
therefore this is going to affect our
future in a 360 degrees all our activity
which we need to process information are
going to change we know that jobs are
going to change and society is going to
change so it’s very important that we
understand what we are talking about in
this context in the case of basic
science for instance you could do
nowadays these techniques are
used for analyzing data that lead to
climate changes they are neat used to
optimize the process of gene editing a
very advanced molecular biology
technique they used to reconstruct the
three-dimensional structure of the of
the brain stantly we still don’t know
and it’s hopefully going to to be to
have a huge impact and it’s even used in
astrophysics so the whole spectrum of of
science is is included and I’m not
mentioning you know all the application
in in human in daily human life this is
all fine however as performant as all
these things may appear I should say
that we are only at a very preliminary
stage there are open problems and
challenges ahead of us for instance
about the functioning of these kinds of
devices we still need to understand how
they how learning can be made faster how
the prediction performance can be
improved and act at a more fundamental
level we need to understand what makes
this kind of devices special compared to
other type of devices so where does the
propensity to learn and generalize come
from these are mathematical question
that I don’t want to trace here but
these are open problems which we all
work then our challenges let me mention
two challenges that are very important
and they’re going to give you a clear
idea of what could be the impact in the
future of this kind of technology the
first challenge is to learn to learn
without supervision so more the vast
majority of the data in the world cannot
be annotated classified by humans and
they contain a lot of information that
cannot even be understood by humans okay
so we need to develop algorithm that are
algorithms that are capable to extract
information to identify patterns to make
prediction without human supervision ok
so this is you know a challenge and all
the research which I have just started
to work on these are working on this and
it’s totally open challenge
another critical objective is these kind
of networks can
predict cannot identify the
relationships between causes and effects
in data so somehow this machine have to
become capable of identifying causation
patterns in in the data in the raw data
when this will become possible then this
tool will provide a lot of information
about modeling systems and making very
complicated prediction and making you
know new tools available as an example
of this I mean you can use the tip
network to recognize images – for speech
recognition you can use it for playing
games and so on and you can use it for
automatic translation you can translate
from a language to another language that
are similar and this translation is just
based on statistics but you cannot use
this kind of tools to create any new
type of sentence because they don’t get
the logical relationship between words
okay just try to set up a conversation
between Siri and Google assistant and
you will see that the conversation
becomes totally banal and boring
immediately as a researcher however all
these kind of open problems are actually
exciting we don’t get depressed when we
find open problems we actually get
excited and and I think they give the
reason is that they clearly show that
there is a huge space for improvement
and for and for application across all
disciplines so there are some
fundamental challenges that have you
know there’s a huge space for
improvement a lot of space for research
and a lot of consequences ahead of us so
what should a student study in order to
be prepared to contribute to this field
this is a question that sometimes occurs
to my mind and actually quite often and
what the answer in my opinion is simple
and I have to say that a lot of
colleagues of mine agree
these there’s no shortcut students need
to know the fundamentals they need to
know math they need to know physics they
need to know computer science they need
to know biology they need to know
chemistry they need to know economics
they didn’t to have all the
methodological tools that allow them to
analyze complex systems they need to
have this background this is the only
way they can actually use the cognitive
power provided by these new tools so the
message for students is actually do not
specialize too soon you have to become
strong from a methodological point of
view be able to model complex system
understand all the tools and then use
this tool or even advance this tool so I
hope that this discussion is somehow
clarified which is the current state of
the art in artificial intelligence the
present of artificial intelligence is
not about thinking machine machines is
about augmenting our cognitive
capabilities through this tool that are
already at work and that have an immense
possibility of making further progress
so as a researcher I would really like
to sit on the other side of the stage
together with the students and the
reason is that there are fantastic and
broad challenges ahead of us and
therefore this is a truly remarkable
time to be a student [Applause]
Please follow and like us: