Translator: Evelyn Montes Auditor: Michele Gianella
Think of the brain of a child of one year.
Do you know how many new connections
They are formed every second among his neurons?
About 700.
In a full day
They make more than 60 million new connections.
And each of these new connections
It represents a new shade, a new concept
that the child has learned through his experience in the world.
Children are learning samples,
and it is to them that look the most modern techniques of artificial intelligence.
At a distance of about 50, 60,
since Alan Turing had theorized for the first time,
you are finally doing the road concept
that if we want to create truly intelligent autonomous systems,
we should not try to infuse them with all our knowledge,
but we have to make them capable of learning
in the most efficient and independent as possible.
The first time I had to deal with artificial intelligence
I was about 15,
and in my free time,
as this may sound a little ‘nerd,
I had created a small program, a small play,
in which a virtual Pierino had to escape from a wolf,
it [ ‘s] to be devoured.
The peculiar thing is that Peter had not planned to escape from the wolf,
but I had given only
the ability to move within a small map,
I gave him a chance to see the wolf,
but above all, I had equipped
the ability to learn from experience
through a simple algorithm of artificial intelligence.
And so, after throwing in this world, turned on the program,
I saw Peter first learn
that staying close to the wolf was not a great idea,
and then even learn to run away as fast as possible
as soon as she saw him in the distance.
And that’s how I fell in love artificial intelligence.
And even today,
who design artificial intelligence systems is my job,
and is the focus of the company that I founded,
still see artificial intelligence systems
learn to make decisions independently
Only through experience on the data that we provide
It is for me something really magical.
Now it’s been almost 20 years
from that moment, from that virtual Pierino,
and today the technology is something else entirely.
In fact,
AI today is among us.
For example, you can drive more safely than humans,
It is able to fight in strategy games
which until a few years ago it seemed the only human rule, absolutely.
It is able to recognize us, we and our friends,
when we upload photos to Facebook.
But above all, it helps us in many areas today,
from finance to medicine.
The technological frontier is represented by Deep Learning,
ie deep learning.
The Deep Learning is based on the construction of virtual brains
with thousands of neurons and millions of connections.
And learn through the reorganization of these connections
after each trial and error making
on the data that we provide during training.
So, during the training phase,
try to make decisions, he takes it, is wrong,
and meanwhile it reorganizes its virtual connections within the brain,
and in doing so he learns.
Just like a child, then.
The problem is that, just like a child becomes an adult,
Also for these AI systems
more knowledge means less flexibility in learning.
So, more knowledge, less flexibility in learning.
And that’s a problem.
And it is a problem that is exemplified in particular
in another advanced technique of artificial intelligence
called reinforcement learning, that is, reinforcement learning.
The Reinforcement Learning
It works through virtual and virtual rewards punishment
which are given to the system depending on the outcome of every decision.
So, for example,
a Reinforcement Learning System is learning to play chess
will be rewarded whenever makes a move
which puts it in an advantageous position with respect to the opponent,
while it is being punished when virtually makes a move not appropriate.
In doing so, hand in hand, playing many matches,
learn the system.
The problem is that what occurs
is that, after a while ‘, after he learned a bit’,
the blind system, faced with similar situations,
to always choose the first road
that had brought him a reward in the past.
So, faced with similar situations, it should be safe. It’s afraid of making mistakes.
And this is a huge problem,
because it is said that the first road that brought us a reward
it is also the road that leads to greater success.
As a result there has invented a trick.
And the trick is, from time to time,
not always, but occasionally,
during system training,
to try to prevent him totally afraid of making mistakes.
So, sometimes, the system is not afraid of making mistakes,
and feels free
to experiment with new roads, new decisions.
In this way,
most of the times it fails miserably, of course.
But every now and discover a new path, a new strategy,
a little ‘creative, a bit’ different,
which eventually leads to a dramatically improved results.
And it is through these techniques, this kind of techniques,
they have in common the fact
the aim to enable artificial intelligence systems
to err more, not less,
During the training phase, during their training,
Today artificial intelligence has achieved success that we know.
And is not it ironic all this?
It is ironic that we humans
we criticize the mistakes of others,
that harshly criticize those trial and error there,
we are also the ones
they are giving the error value
artificial intelligence?
We are giving the error value to artificial intelligence,
when we could learn so much from this.
For example, in the Deep Learning
we try in every way to prevent the system
to store data that we provide during training,
because if they did,
totally lose the flexibility to adapt to new situations.
So why do we humans
we force such as our children
to employ a lot of time learning to date and data storage
that inevitably forget to first check?
Because the education system, often,
It is based more, and more rewards
saving compared to an understanding?
And yet, in Reinforcement Learning
we have seen that fear of failure is seen as a problem,
for which you search every so often to allow, during the formation,
not to be afraid to make mistakes
so that the system can find out the roads leading to the then greatest hits.
So why, between us,
when someone changes the road during their training,
or even during his career,
It is seen as a total reckless?
Because who changes his mind,
as a result of life experiences, as a result of reflections,
It is first seen as inconsistent
that wise?
And finally,
we have seen and known,
that artificial intelligence learns something
all the times wrong.
So why do we continue to see the error as a defeat,
when actually it is only the first step
on the road of change and success?
Thank you.
(Applause)
Think of the brain of a child of one year.
Do you know how many new connections
They are formed every second among his neurons?
About 700.
In a full day
They make more than 60 million new connections.
And each of these new connections
It represents a new shade, a new concept
that the child has learned through his experience in the world.
Children are learning samples,
and it is to them that look the most modern techniques of artificial intelligence.
At a distance of about 50, 60,
since Alan Turing had theorized for the first time,
you are finally doing the road concept
that if we want to create truly intelligent autonomous systems,
we should not try to infuse them with all our knowledge,
but we have to make them capable of learning
in the most efficient and independent as possible.
The first time I had to deal with artificial intelligence
I was about 15,
and in my free time,
as this may sound a little ‘nerd,
I had created a small program, a small play,
in which a virtual Pierino had to escape from a wolf,
it [ ‘s] to be devoured.
The peculiar thing is that Peter had not planned to escape from the wolf,
but I had given only
the ability to move within a small map,
I gave him a chance to see the wolf,
but above all, I had equipped
the ability to learn from experience
through a simple algorithm of artificial intelligence.
And so, after throwing in this world, turned on the program,
I saw Peter first learn
that staying close to the wolf was not a great idea,
and then even learn to run away as fast as possible
as soon as she saw him in the distance.
And that’s how I fell in love artificial intelligence.
And even today,
who design artificial intelligence systems is my job,
and is the focus of the company that I founded,
still see artificial intelligence systems
learn to make decisions independently
Only through experience on the data that we provide
It is for me something really magical.
Now it’s been almost 20 years
from that moment, from that virtual Pierino,
and today the technology is something else entirely.
In fact,
AI today is among us.
For example, you can drive more safely than humans,
It is able to fight in strategy games
which until a few years ago it seemed the only human rule, absolutely.
It is able to recognize us, we and our friends,
when we upload photos to Facebook.
But above all, it helps us in many areas today,
from finance to medicine.
The technological frontier is represented by Deep Learning,
ie deep learning.
The Deep Learning is based on the construction of virtual brains
with thousands of neurons and millions of connections.
And learn through the reorganization of these connections
after each trial and error making
on the data that we provide during training.
So, during the training phase,
try to make decisions, he takes it, is wrong,
and meanwhile it reorganizes its virtual connections within the brain,
and in doing so he learns.
Just like a child, then.
The problem is that, just like a child becomes an adult,
Also for these AI systems
more knowledge means less flexibility in learning.
So, more knowledge, less flexibility in learning.
And that’s a problem.
And it is a problem that is exemplified in particular
in another advanced technique of artificial intelligence
called reinforcement learning, that is, reinforcement learning.
The Reinforcement Learning
It works through virtual and virtual rewards punishment
which are given to the system depending on the outcome of every decision.
So, for example,
a Reinforcement Learning System is learning to play chess
will be rewarded whenever makes a move
which puts it in an advantageous position with respect to the opponent,
while it is being punished when virtually makes a move not appropriate.
In doing so, hand in hand, playing many matches,
learn the system.
The problem is that what occurs
is that, after a while ‘, after he learned a bit’,
the blind system, faced with similar situations,
to always choose the first road
that had brought him a reward in the past.
So, faced with similar situations, it should be safe. It’s afraid of making mistakes.
And this is a huge problem,
because it is said that the first road that brought us a reward
it is also the road that leads to greater success.
As a result there has invented a trick.
And the trick is, from time to time,
not always, but occasionally,
during system training,
to try to prevent him totally afraid of making mistakes.
So, sometimes, the system is not afraid of making mistakes,
and feels free
to experiment with new roads, new decisions.
In this way,
most of the times it fails miserably, of course.
But every now and discover a new path, a new strategy,
a little ‘creative, a bit’ different,
which eventually leads to a dramatically improved results.
And it is through these techniques, this kind of techniques,
they have in common the fact
the aim to enable artificial intelligence systems
to err more, not less,
During the training phase, during their training,
Today artificial intelligence has achieved success that we know.
And is not it ironic all this?
It is ironic that we humans
we criticize the mistakes of others,
that harshly criticize those trial and error there,
we are also the ones
they are giving the error value
artificial intelligence?
We are giving the error value to artificial intelligence,
when we could learn so much from this.
For example, in the Deep Learning
we try in every way to prevent the system
to store data that we provide during training,
because if they did,
totally lose the flexibility to adapt to new situations.
So why do we humans
we force such as our children
to employ a lot of time learning to date and data storage
that inevitably forget to first check?
Because the education system, often,
It is based more, and more rewards
saving compared to an understanding?
And yet, in Reinforcement Learning
we have seen that fear of failure is seen as a problem,
for which you search every so often to allow, during the formation,
not to be afraid to make mistakes
so that the system can find out the roads leading to the then greatest hits.
So why, between us,
when someone changes the road during their training,
or even during his career,
It is seen as a total reckless?
Because who changes his mind,
as a result of life experiences, as a result of reflections,
It is first seen as inconsistent
that wise?
And finally,
we have seen and known,
that artificial intelligence learns something
all the times wrong.
So why do we continue to see the error as a defeat,
when actually it is only the first step
on the road of change and success?
Thank you.
(Applause)
Please follow and like us: