Jeremy Howard: The wonderful and terrifying implications of computers that can learn
Jeremy Howard - Data scientist
Jeremy Howard imagines how advanced machine learning can improve our lives. Full bio
to get a computer to do something new,
that haven't done it yourself,
the computer to do
that you don't know how to do yourself,
to be a great challenge.
by this man, Arthur Samuel.
how to be better than you at checkers?
against itself thousands of times
and in fact, by 1962,
the Connecticut state champion.
the father of machine learning,
machine learning practictioners.
previously unsolved problems,
hundreds of times.
I was able to find out
can do in the past, can do today,
machine learning commercially was Google.
possible to find information
on machine learning.
commercial successes of machine learning.
products that you might like to buy,
who your friends might be
the power of machine learning.
learned how to do this from data
the two world champions at "Jeopardy,"
and complex questions like this one.
from this city's national museum in 2003
(along with a lot of other stuff)"]
to see the first self-driving cars.
the difference between, say,
well, that's pretty important.
those programs by hand,
this is now possible.
over a million miles
don't know how to do ourselves,
I've seen of machine learning
called Geoffrey Hinton
automatic drug discovery.
is not just that they beat
or the international academic community,
in chemistry or biology or life sciences,
called deep learning.
the success was covered
article a few weeks later.
here on the left-hand side.
inspired by how the human brain works,
on what it can do.
computation time you give it,
showed in this article
result of deep learning
can listen and understand.
to take in this process
of information from many Chinese speakers
and converts it into Chinese language,
an hour or so of my own voice
so that it would sound like me.
a machine learning conference in China.
at academic conferences
at TEDx conferences, feel free.
was happening with deep learning.
was deep learning.
in the top right, deep learning,
was deep learning as well.
this extraordinary thing.
can seem to do almost anything,
it had also learned to see.
to recognize traffic signs like this one.
recognize the traffic signs
it was better than people,
better than people.
they had a deep learning algorithm
on 16,000 computers for a month,
about concepts such as people and cats
that humans learn.
by being told what they see,
what these things are.
who we saw earlier,
from one and a half million images
to a six percent error rate
an extraordinarily good job of this,
location in France in two hours,
that they fed street view images
to recognize and read street numbers.
it would have taken before:
the Chinese Google, I guess,
to Baidu's deep learning system,
has understood what that picture is
have similar backgrounds,
at the text of a web page.
really understand what they see
of images in real time.
now that computers can see?
that computers can see.
has done more than that.
with deep learning algorithms.
showing the red dot at the top
is expressing negative sentiment.
is near human performance
and what it is saying about those things.
been used to read Chinese,
Chinese speaker level.
out of Switzerland
or understand any Chinese.
in the world for this,
put together at my company
all this stuff together.
have no text attached,
to the text that I'm writing.
understanding my sentences
something like this on Google,
and it will show you pictures,
searching the webpage for the text.
understanding the images.
have only been able to do
can not only see but they can also read,
can understand what they hear.
I'm going to tell you they can write.
using a deep learning algorithm yesterday.
out of Stanford generated.
to describe each of those pictures.
a man in a black shirt playing a guitar.
it's seen black before,
this novel description of this picture.
performance here, but we're close.
the computer-generated caption
well past human performance
to very exciting opportunities.
that they had discovered
make a prognosis of a cancer.
looking at tissues under magnification,
a machine learning-based system
than human pathologists
for cancer sufferers.
were the predictions more accurate,
that humans can understand.
that the cells around the cancer
the cancer cells themselves
had been taught for decades.
they were systems developed
and machine learning experts,
we're now beyond that too.
identifying cancerous areas
can identify those areas more accurately,
as human pathologists,
using no medical expertise
no background in the field.
about as accurately as humans can,
with deep learning
background in medicine.
no previous background in medicine,
to start a new medical company,
that it ought to be possible
using just these data analytic techniques.
has been fantastic,
but from the medical community,
the middle part of the medical process
as much as possible,
what they're best at.
to generate a new medical diagnostic test
three minutes by cutting some pieces out.
creating a medical diagnostic test,
a diagnostic test of car images,
we can all understand.
about 1.5 million car images,
that can split them into the angle
so I have to start from scratch.
areas of structure in these images.
and the computer can now work together.
about areas of interest
to try and use to improve its algorithm.
are in 16,000-dimensional space,
rotating this through that space,
point out the areas that are interesting.
successfully found areas,
the computer more and more
we're looking for.
areas of pathosis, for example,
potentially troublesome nodules.
difficult for the algorithm.
of the cars are all mixed up.
as opposed to the backs,
that this is a type of group
we skip over a little bit,
machine learning algorithm
some of these pictures out,
how to understand some of these itself.
of similar images,
entirely find just the fronts of cars.
can tell the computer,
a good job of that.
to separate out groups.
computer try to rotate this for a while,
and the right sides pictures
the computer some hints,
a projection that separates out
as much as possible
ah, okay, it's been successful.
of thinking about these objects
is being replaced by a computer,
something that used to take a team
that takes 15 minutes
four or five iterations.
can start to quite quickly
that there's no mistakes.
let the computer know about them.
for each of the different groups,
an 80 percent success rate
that aren't classified correctly,
to 97 percent classification rates.
could allow us to fix a major problem,
of medical expertise in the world.
that there's between a 10x and a 20x
in the developing world,
to fix that problem.
enhance their efficiency
about the opportunities.
every area in blue on this map
are over 80 percent of employment.
computers have just learned how to do.
in the developed world
have just learned how to do.
They'll be replaced by other jobs.
more jobs for data scientists.
very long to build these things.
were all built by the same guy.
it's all happened before,
of when new things come along
grows at this gradual rate,
in capability exponentially.
are still pretty dumb." Right?
computers will be off this chart.
about this capability right now.
in capability thanks to engines.
that after a while, things flattened out.
to generate power in all the situations,
from the Industrial Revolution,
it never settles down.
at intellectual activities,
to be better at intellectual capabilities,
never experienced before,
of what's possible is different.
as capital productivity has increased,
in fact even a little bit down.
having this discussion now.
about this situation,
they don't understand poetry,
of their time being paid to do,
social structures and economic structures
About the speaker:Jeremy Howard - Data scientist
Jeremy Howard imagines how advanced machine learning can improve our lives.
Why you should listen
Jeremy Howard is the CEO of Enlitic, an advanced machine learning company in San Francisco. Previously, he was the president and chief scientist at Kaggle, a community and competition platform of over 200,000 data scientists. Howard is a faculty member at Singularity University, where he teaches data science. He is also a Young Global Leader with the World Economic Forum, and spoke at the World Economic Forum Annual Meeting 2014 on "Jobs for the Machines."
Howard advised Khosla Ventures as their Data Strategist, identifying opportunities for investing in data-driven startups and mentoring portfolio companies to build data-driven businesses. He was the founding CEO of two successful Australian startups, FastMail and Optimal Decisions Group.
Jeremy Howard | Speaker | TED.com