Principles For Human-Centered AI | Michael I Jordan (UC Berkeley)
Summary
TLDRThe speaker advocates for a market-driven approach to AI, emphasizing decision-making and economic principles over human imitation. They argue that intelligent systems should create value by connecting producers and consumers, like Amazon and Uber, rather than just mimicking human intelligence. The talk highlights the potential of machine learning in creating robust, enjoyable human experiences by focusing on market mechanisms and data-driven decision-making.
Takeaways
- 🚀 The future of business models should focus on becoming more like Netflix, Uber, or Amazon by leveraging intelligent systems and market mechanisms.
- 🤔 The term 'intelligent systems' is often used without a clear definition, prompting the need to look at Earth from a Martian perspective to identify what is truly intelligent.
- 🧠 Traditional AI has focused on mimicking human intelligence, but this may not be the most effective approach for creating robust and beneficial systems.
- 🏙️ Cities and markets demonstrate a form of intelligence through their ability to coordinate complex systems for the distribution of goods and services.
- 🛠️ The new frontier in computer science is not just about what happens inside the computer, but also how computers interact with the world through market mechanisms.
- 🚗 The concept of 'autonomous' vehicles is flawed; instead, vehicles should interact with a larger system, similar to air traffic control, for better coordination.
- 🔢 Machine learning has evolved from backend data analysis to customer-facing recommendation systems, and now to pattern recognition and decision-making.
- 💡 The speaker advocates for a shift in AI research from human imitation to creating systems that make better decisions in the context of markets and scarcity.
- 💼 Businesspeople should consider machine learning as having two parts: pattern recognition and decision-making, with the latter being crucial for creating value.
- 🎵 The music industry is an example where a market exists but is not fully monetized, suggesting opportunities for new business models that connect producers and consumers.
- 🌐 The potential for AI is vast, extending beyond traditional areas like music to art, cooking, and personal services, where connecting providers with consumers can create jobs and wealth.
Q & A
What is the main focus of the talk given in the transcript?
-The talk focuses on the business model of the future, discussing how to become a company like Netflix, Uber, or Amazon, and the importance of technical fields driving business strategic level thinking.
Why does the speaker suggest considering oneself as a Martian computer scientist?
-The speaker suggests this perspective to encourage out-of-the-box thinking and to look for inspiration for intelligent systems on Earth that could be mimicked or improved upon in computer algorithms.
What is the speaker's view on the current approach to artificial intelligence and its focus on imitating human intelligence?
-The speaker believes that the focus on imitating human intelligence is misguided and that a more promising approach involves creating market mechanisms and systems that interact with the world in a more integrated way.
What does the speaker consider as an example of intelligence on Earth that could inspire computer algorithms?
-The speaker considers the market systems, like the coordination of ingredients and supplies in a city's restaurants, as an example of intelligence that operates efficiently and could inspire computer algorithms.
Why does the speaker argue that the term 'autonomy' in the context of self-driving cars is incorrect?
-The speaker argues that 'autonomy' is incorrect because self-driving cars should not operate in isolation like humans but should interact with all other cars in a more integrated system, similar to air traffic control.
What is the speaker's opinion on the current state of machine learning and AI in terms of economic impact?
-The speaker believes that while machine learning and AI have had significant impacts, especially in backend operations and recommendation systems, the next generation of AI should focus more on decision-making and market mechanisms for greater economic value.
What is the concept of 'two-way markets' as mentioned by the speaker?
-The concept of 'two-way markets' refers to a system where both sides of a transaction can see and bid on each other, creating a more efficient and dynamic market that reduces noise and increases the likelihood of successful transactions.
How does the speaker suggest using recommendation systems in a new way to create value?
-The speaker suggests using recommendation systems not just to suggest items but to facilitate real-time bidding and transactions, such as recommending a customer to a restaurant and allowing the restaurant to offer a discount in return.
What is the speaker's view on the potential of market-based AI and its impact on jobs and the economy?
-The speaker believes that market-based AI has the potential to create jobs and generate wealth by connecting producers and consumers directly, rather than relying on advertising or other indirect monetization methods.
What are some of the technical problems the speaker is working on to advance market-based AI?
-The speaker is working on problems such as multiple decisions, markets under uncertainty, cloud-edge interactions, and provenance, which are essential for powering the next generation of market-based AI.
Can you provide an example of how the speaker suggests applying market-based AI to the music industry?
-The speaker suggests creating a platform that not only streams music but also provides data to musicians about their popularity in different cities, enabling them to negotiate performances and create a sustainable income, thus fixing the broken music market.
Outlines
🚀 The Future Business Model and AI's Role
The speaker addresses CEOs, discussing the future of business models, inspired by companies like Netflix, Uber, and Amazon. They question the common understanding of 'intelligence' in AI, urging the audience to consider alternative forms of intelligence observed in systems like city markets, which are robust, reliable, and adaptable. The speaker suggests that the real innovation lies in market mechanisms and the interaction of computers with the world, rather than just imitating human intelligence. The talk emphasizes the importance of strategic thinking at the intersection of technology and business.
🤖 Rethinking AI: Beyond Human Imitation
The speaker critiques the current focus on imitating human intelligence in AI, arguing for a broader perspective that includes decision-making and market dynamics. They discuss the evolution of machine learning from backend processes to customer-facing recommendation systems, which have significantly impacted industries. The speaker advocates for considering machine learning as having two parts: pattern recognition and decision-making, with the latter being crucial for real-world applications and economic value generation.
🌐 Creating Two-Way Markets with AI
The speaker introduces the concept of using AI to create two-way markets, exemplified by a recommendation system for restaurant seats. They propose a system where customers' preferences are broadcasted to nearby restaurants, allowing them to bid for the customer's business, thus creating a dynamic market that reduces noise and increases efficiency. The speaker suggests that this approach can be applied to various economic goods and services, creating new business opportunities and revenue streams through transaction fees.
💡 Empowering People with Market Mechanisms
The speaker argues for a shift from trying to understand and predict human desires through data to empowering people through market mechanisms. They propose a model where data flows lead to economic value and transactions, creating happier customers and a more honest way of making a profit through transaction fees. The speaker also touches on the social benefits of creating jobs by connecting people and providing services, suggesting that this approach is the next big opportunity for businesses.
🎵 The Music Industry as a Market Model
Using the music industry as an example, the speaker illustrates how a market-based approach can revitalize an industry where traditional models have failed. They propose a system where musicians can see data on their popularity in different cities and use this information to negotiate performances and other opportunities, thus creating a sustainable income. The speaker suggests that this model can be applied to various domains, emphasizing the potential for wealth creation and job generation.
🔬 Research Directions for Market-Based AI
The speaker concludes with a discussion on research directions for developing market-based AI systems. They highlight the importance of addressing multiple decisions, markets, uncertainty, and the interaction between cloud and edge computing. The speaker also mentions the need for algorithms that can handle competition, such as bandit algorithms in the presence of multiple decision-makers, and the importance of controlling false discovery rates in large-scale A/B testing.
Mindmap
Keywords
💡Intelligent systems
💡Markets
💡Autonomy
💡Machine learning
💡Recommendation systems
💡Economic perspective
💡Scarcity
💡False discovery rate
💡Two-way market
💡Gig economy
Highlights
The talk focuses on the business model of the future, aiming to inspire CEOs to think like Netflix, Uber, or Amazon.
Intelligent systems are often misunderstood; the speaker suggests viewing them from a 'Martian' perspective for fresh insights.
The speaker argues that human brains, while intelligent, may not be the best model for Martian-inspired computer architecture due to their Earth-specific adaptations.
Cities and their complex systems, like San Francisco's restaurant supply chains, are presented as examples of distributed intelligence and robustness.
Markets are identified as inherently intelligent systems that have been driving human happiness and viability for thousands of years.
The importance of microeconomics and game theory in understanding market intelligence is emphasized.
The speaker criticizes the current focus on imitating human intelligence in AI, suggesting a shift towards creating systems that interact with the world.
The concept of 'autonomy' in self-driving cars is challenged, proposing a more integrated system similar to air traffic control.
The evolution of machine learning from backend operations to customer-facing recommendation systems is outlined.
The potential of machine learning to create new billion-dollar industries through pattern recognition is discussed, but with skepticism on its economic value.
The speaker introduces the idea of machine learning having two parts: pattern recognition and decision making, with a focus on the latter.
The importance of considering the consequences of decisions made by machine learning systems is highlighted.
Recommendation systems are reimagined not just as tools for suggesting products, but as mechanisms for creating two-way markets.
The potential of using recommendation systems to create dynamic markets in various industries, such as music, art, and cooking, is explored.
The speaker proposes a new formula for AI involving data, algorithms, and markets, moving away from human imitation.
The social and economic benefits of creating jobs and wealth through market-based AI are discussed.
Examples of companies like Amazon that understand the importance of connecting producers and consumers are given.
The speaker shares his research interests in areas such as multiple decisions, markets, uncertainty, and cloud-edge interactions.
Innovative projects in bandit algorithms and asynchronous online false discovery rate control are mentioned, showing the future of market-based AI.
Transcripts
all right I'm pleased to be here so this
talk is a little different than the one
I usually give this is a talk not aimed
at the mathematicians in the audience so
much as the CEOs so what I'm going to be
trying to convey is what I consider the
business model of the future how do you
become a Netflix or an uber or an Amazon
what's missing out there so how is it as
our technical field drive business
strategic level thinking so we all claim
that we're working on intelligent
systems it would be great if computers
were more intelligent whatever that
means but that's kind of the issue what
does that mean we just use that word as
if we all know what we're talking about
so extract yourself from the earth for a
minute go up to Mars and imagine you're
a Martian computer scientist and you
have a very primitive computer
architecture and you're looking down at
earth to try to get inspiration for how
to make it better all right so you look
down and say what's intelligent on the
planet earth that I could try to mimic
and put inside my algorithms and my
computers so the first thing you might
notice is these things brains and minds
and you might agree that there are some
how intelligent whatever it means on the
other hand you might note that they're
pretty well adapted to earth and maybe
the intelligence that are exhibiting is
pretty adapted to earth and not so
adapted to Mars or to other situations
moreover understanding brains and nert
webs of neurons and how that leads to
thought and behavior seems really
challenging and in fact I argue that
even on planet Earth where we get this
look at and probe at a quick pretty
closely we're very very far away from
understanding in fact we don't have
really a glimmering yet of how thought
arises from webs of neurons all right so
as the martian computer science just
kind of lost what else down on earth is
intelligent that you could algorithm a
size and all right well it's not that
hard if you just think out of the box a
little bit and I don't think enough
people are thinking out of the box so
think about a city like San Francisco
and think about every restaurant
and think about every dish that the
restaurants creating it has ingredients
and those ingredients have got to arrive
at the restaurant every day and it
happens there's a web of decisions being
executed distributed slightly
coordinated worth individuals deciding
to bring you know their goods from one
side of the city the next and so on and
it works all those restaurants get what
they need and every home gets what it
needs more or less and it doesn't work
perfectly that's always interesting but
it works really well works at all scales
small big scale City it works for 3,000
years
okay and it's probably done more to make
human happiness and life viable than
anything else in the world so it's
markets markets are intelligent by any
definition they're robust they're
reliable they work in all conditions
rain or shine day in and day out and we
know some of the principles it's called
microeconomics okay it's game theory
it's trade its markets and so on we
don't know all the principles but that's
what it's interesting about it
what new principles are needed if we
bring that into the world of the
computer I know what does it mean to
bring in the world of computer well if
it was just inside the computer that's
maybe not so interesting that's like
resource allocation or whatever fine but
we're no longer just inside the computer
that's what's new about computer science
is that it's out in the world and in the
computer in the bridge is the hard part
and if you just still think about the
old computer science mentality and build
a service in the computer and then make
it so good that people are attracted to
it like you know moths to a flame and
then understand everything about the
people just by being inside the computer
that's kind of the world we're in and
that's kind of not working very well
you've got to have market mechanisms
where the computers interacting with the
world in the sense of an actual market
alright there's another way to say this
people have been too focused on this
notion of imitating human intelligence
that's what AI referred to in the 50s
and now that we brought that terminology
back to describe machine learning people
think that's what we're still working on
is imitating humans so that we can take
out a driver in a car and put in a
computer just like the human or a
dialogue person we put in a human a
computer right no that's not the really
this kind of intelligence we can it she
in our lifetime that will make human
life more robust more fun more pleasant
and so on
think about autonomous driving cars that
word autonomy is just dead wrong
you don't want self-driving cars to be
like a human
that's autonomous and you put it in the
car in the car is suddenly autonomous
right no if that car should interact
with all the other cars around it in a
much deeper way than we currently do
right if one car knows that boys just
ran out into the street every other car
around that should know immediately okay
it should be more like the air traffic
control system it should be an
integrated system okay autonomy is just
the wrong way to think about the whole
thing all right and even intelligence is
not embedded in the car itself and
intelligence is better in the entire
system right I've given this slide
before I think in fact it's SPARC last
year so I won't spend too much time on
it but just look at a little bit of the
history of this field that I called
machine learning for many years that
people are now calling AI you know it's
not new already in the 90s the backend
was extremely important a fraud
detection search supply chain management
so a company like Amazon would have
never become Amazon without doing
machine learning in the backend at very
big scale with big data and lots of
computers and lots of Engineers they did
that and they became Amazon I think the
next generation was noticing that all
that data analysis could be turned not
just at the back end but turned towards
the customer and recommendation systems
emerge so you saw a great talk about
that just now that was already the
second generation and then the Netflix
and Amazon kind of led that those two
generations led to billion-dollar
industries
okay no one talks about it certainly the
journalists don't talk about it but that
had a huge impact on our world this
third generation which I think it was
just really pattern recognition finding
patterns and data you know the success
stories are things like speech
recognition of vision oh yeah I don't
see the economic value yet
okay now many of you're working on that
it'll be great
but I don't see the billion dollar
industry yet okay because you're kind of
just taking out a human and putting in a
computer and it's not so third of me
you're gonna you know make the world a
better place or you're gonna make a lot
of money doing that
okay not so clear and in fact I think
what's really emerging is this web of
decisions these markets that a computer
with data analysis can create if you
think in that way and that's what I want
you to think about as a businessperson
all right so instead of thinking about
pattern recognition is the big
achievement and the big technology
tensorflow if you will think about
machine learnings having two parts to it
pattern recognition and decision making
all right now decision making what is
that well isn't that just taking the
output of tensorflow and threshold in it
okay well no all right so think about
things that have consequences so I go
into the doctor's office and you measure
a lot of things about me my DNA
my blood pressure and all that you know
hundreds of thousands of variables you
put that into a big neural net and it
says you have liver disease you need to
have your liver transplant it tomorrow
urgently what does it mean it says that
well there's some number that's over
some threshold point seven and it's 0.71
all right is that a decision okay well
no I'm not gonna just stop there I'm
gonna say wait a minute what data are
you using to make that prediction is
that data recent is it about people like
me what if I were to exercise more or
what if I did this or what I did you
know this thing about me or about my
past or whatever etcetera so I want a
whole dialogue and that's just for one
decision okay but one that has
consequences real-life decisions have
consequences we want to talk about
Providence where the data came from
counterfactuals and have a whole
dialogue all right but decisions are
never in isolation we're always making
decisions in the context of other
people's decisions other agents there's
going to be scarcity we also have to
worry about false discovery rate if I'm
gonna have a system make a hundred
thousand decisions today like uber who's
allocating cars to places in the city
right I want most of that Baguio
decisions to be good ones if I think
about them independently that's no good
I want the overall set of them to be
good okay moreover I really want to
think about sets of decisions by sets of
agents over time and I want to see as
time goes on your air rate is not kind
of wildly fluctuating it's it's
controlled in some nice way and I want
asynchronous AV testing is kind of a
I'm aware we use data to make good
decisions over time by the large set
collections of agents who are loosely
coupled that's the real-world problem to
me it's not just taking an image of a
computer you know on a computer or
taking and saying there's a kangaroo in
the image decision okay all right and
even taking it further when we have
competition and scarcity we really need
to have more than just making decisions
with thresholds and reinforcement
learning and all that we need an
economic perspective okay okay so let's
look at recommendation systems which you
heard about in the previous talk from an
expert you all know what they are you
keep a record of customers purchases if
customers were similar you recommend
items from one customer to another
customer and very important you know
let's call it a billion-dollar industry
most companies are doing this it's
become a commodity you can download
software to do it and so a lot of people
look at that and say great I can build a
company because this off the platform
exists I can get the computers I can get
the image I can download I can run a
recommendation system for something else
than movies all right and build a
company all right if you do that you're
gonna run into all kinds of trouble
because you're not thinking out of the
box you're thinking about the old model
recommending movies or books or whatever
right what if I recommend the same movie
to everybody well everybody's happy how
will you stream the bits to everybody
else care city same thing with books
nowadays you can print it on demand the
Amazon can get it to you within three
days even if they recommend the same
book to everybody right but now if you
build a system that recommends other
kind of things that are economic goods
like restaurant seats I come out on out
of the you know a taxi in Shanghai at
the end of a datum they have meetings I
don't know Shanghai very well I don't
speak Mandarin I'm by myself I want to
have a recommendation for a place to go
eat okay but not just a recommendation I
just a list or a set of advertisements
or Yelp pages or something like that I
want to have a button where I push it
and I say geo locate me now broadcast to
all the restaurants around me that I'm
here
make it a recommendation system meaning
that you see something about me that I
have a certain price point
I liked Szechuan cuisine etc etc let the
restaurants see that and let them bid on
me that restaurant says we like you
the next 10 minutes will give you a 10%
discount if you push the button we just
created a two-way market and both sides
are seeing the other through a
recommendation system all right so
that's gonna cut down all the noise all
right I'm not gonna see 500,000
restaurants with Yelp reviews and some
on the other side they're gonna see
customers that are viable and as soon as
we make a transaction we fill a seat in
the restaurant you're not going to
recommend that same seat to a hundred
thousand people okay simple but very
important now the company that does this
has just built a two-way market how do
they make their money well it's way
easier than build an advertising Empire
rather you take 5% of your cut of the
transactions all right and you can make
big money that way if this is really
rolled out throughout Shanghai you're
gonna make tons of money and then
worldwide you know great what if I were
to recommend streets to drivers
everybody needs to get to the airport if
I recommend the same Street to everybody
it's no longer a good Street now this is
obvious right but people don't tend to
think this way they roll out their
service and then maybe a thousand people
are using it and everybody's happy man
the whole city starts to use it and
everybody's not happy anymore oh oh you
should have thought of that right but
people are not doing it now why are they
not doing it well because they have this
AI mentality all right let's just take
the kind of Facebook example they think
that I'm gonna figure out what you want
by looking at your browsing history so
if I'm going to the airport today
all right there's several options maybe
I'm really in a rush and I really want
to go on the fastest street and I'll pay
a little bit more for that all right
whereas you're going to the airport and
you think about it for me and say hey
I'm not in such a rush I'll go on a
slightly slower Street and I'll save
some money for some pre next day that's
a little market kind of mechanism right
you could build something like that and
people could have a little interface
where they could start to start to say
and everybody be happier the whole thing
would flow faster right Silicon Valley
instead is thinking in the following way
we're gonna figure out so we're gonna do
a big dynamic program and figure out
where each person wants to go and then
we're gonna load balance and decide who
goes down what Street how are we gonna
figure that out well we know your
browsing history what you know
that's that's insane but that's where
we're at people think that you're gonna
see everything about us and then offer
us the best things that we really really
want even don't even know it all right
that mentality is coming from too much
of a wedding of advertising and company
all right
instead we want to empower people to be
connected to each other producers and
consumers uber does this so it's not a
mystery you can do this writers and
drivers does uber have to advertise to
make money you know no they can take a
transaction fee all right so diners are
on one side of the market restaurants
are on the other you don't have to do
any advertising at all you just connect
them and then you take a cut drivers are
on one side of the Market Street
segments are on the other the street
segment bids for me to go across the
street all right so it's not just
classical microeconomics that I'm
arguing for here that is part of it but
it's also at scale and it's a
recommendation system mentality that we
have a lot of data that doesn't tell us
what people want directly all right but
it gives us some honing down so they can
have a straight ability simple market in
which they participate in and one that's
adaptive okay so if we're gonna have a
formula let's call AI data plus
algorithms plus markets critically and
let's leave behind the human imitative
side of this it's fine if you figure out
how the brain works great and build a
company on that I just not gonna I'm not
gonna invest but if you think you've got
it fine
but I think if you build market
mechanisms I'm ready to invest because
there's so much sitting there on that
that's not being monetized in any way so
wherever you have data flows instead of
thinking of that as just how do I get
the data to flow faster and how do i
monetize that with advert or with
something or advertising think about if
data is flowing probably some economic
value is flowing on the on top of that
and if I really connect up on both sides
and make transactions appear on top of
that probably everyone's going to be
happier really like my service not be so
annoyed about privacy issues because
they're getting value out of it and
probably I can take a transaction fee
which is a perfectly honest way to make
a living okay so for AI research this
autonomy and human imitative side I
think is his is needs to be diminished
we need to think more about federated
agents that interact with each other in
the world of scarcity IT business
instead of building a service that tries
to figure out everything about humans
and give them what you think they want
and make money from an artificial market
between you and advertisers directly
connect people to each other all right
so the social consequences of this
you're gonna be able to create jobs when
you connect people to each other you've
essentially created a job the provider
for provides something to somebody and
they pay that's the job all right
and now we can what kind of domains can
we do this in and this is where I'm
getting closer to the business model all
right so I'm gonna give an example which
I like to give it's music we all think
that music is like hot you know it's
everywhere you know it's being streamed
and everything where they hear music
well know anybody who's a musician I'll
tell you they cannot make money being a
musician Beyonce can but that's um
that's not a market that's a monopoly
all right all right there are tons of
people and probably many of you do it on
the weekend you go home and you make
music on your laptop and you uploaded to
SoundCloud a lot of people drive taxis
in the week and they do that on the
weekend and the music's really good if
you go to SoundCloud it's so good that
companies like Spotify and all streaming
it to people and tons of people listen
to it in fact most music me listen to in
this very moment is music done in the
last four years and it's done by people
who just put it up on soundcloud it's
not Beyonce at all so somehow there's a
wonderful market city there but it's not
a market no one's making money off of it
now we don't have a job doing that all
right well how can you fix this instead
of just creating a streaming platform
for music and then monetize it by
advertising well it's not that hard it's
really easy be a data scientist so
anybody who's putting music on
soundcloud it may be over some threshold
of popularity at the end of the week
gets to see a dashboard alright and that
dashboard has that on every city in
whatever country you're in and you see
how many people listen to you in each
city alright so if my musics popular
maybe and you know Cleveland Ohio 10,000
people listen to me last week you know I
didn't know that now that I know that
I'm gonna go to the venue owners and
Cleveland show them that data and then
they're gonna say wow you're popular
here can you give a show here I'll say
yeah I fly out there I give a show I'm a
twenty thousand dollars if I do that
three times during the year I have a
salary I could quit my taxi job which is
about ready to disappear
anyway and I could be a musician and
moreover now that I know who listens to
me and comes to my show that's easy too
you know I just have a little QR code on
the screen while giving the show I can
make them other offers I'll say I'll
come plate your daughter's wedding for
you know $20,000 $10,000 we'll have a
little negotiation right that's a market
that's gonna create so much wealth and
then the company providing all this
right this is not just a little
peer-to-peer thing the company has to
provide this and do it well takes a cut
of all that okay rich you can make me a
lot of money doing this all right that's
just for music think about art thinking
about cooking thinking about other
personal services of all kinds so this
is kind of the share of the gig economy
if you will all right but I don't think
people are getting how big this really
is this is the next trillion-dollar
thing to do this I think that Amazon
gets this and I think that a lot of what
you see all the innovation from Amazon
right now
okay it's not just because they're so
damn smart it's because they see that
they're trying to provide services to
people and connect producers and
consumers they get that all right I
don't think Facebook gets it they think
they're supposed to just connect people
and make a community of whatever that
means
okay
so I hope that inspires some of you to
think about market metals in your own
business now I'm a researcher so well
first of all I do want to mention
there's a company already doing this
they beat you to it it's called United
masters go online so there's some famous
musicians behind this who themselves
make tons of money but they don't like
the system and they think it's broken
and they want to fix it so they're
paired with computer scientists to build
something called United masters but
there's still funds tons of opportunity
that net1 discreetly the company that
survives okay now I'm a researcher I
work on technical problems in this area
this is a list that I've been using for
the last few years kind of summarizing
some of them you know I work on multiple
decisions I work on markets uncertainty
cloud edge interactions abstractions
provenance and all this
it sounds like boring back-end stuff if
you will or computer science II kind of
stuff there's no sexy AI language on
there but to me these are the problems
to solve and this is good to power the
next generation of market-based AI and
so I have slides which I'm going to kind
of skip a little bit but I just want to
tell you I 30 seconds left here here's
some projects if you'll go to my website
you'll see some of the emerging work on
this topic we've been working on banded
algorithms bandits are a really great
thing to think about your machine
learning person it's not reinforcement
learning it's not supervised learning is
kind of in-between it's that you have
several options and no one's telling you
which is the best option so you have to
try them out a little bit so a be tests
have this flavor so we're doing that
when there's competition what if I have
multiple people doing a bandit algorithm
and there's competition only one of them
can have the arm that they if they all
select the same arm only one of them
gets the arm okay brand new problem
obviously big implications for real
world problems finding Nash equally with
Grady based algorithms and high
dimensional accent spaces this blends
machine learning style thinking with
microeconomics thinking and then
asynchronous online false discovery rate
control if you're doing in large-scale
ad testing in any company you want to
have false discovery rate control you
want it to be asynchronous and you want
to be online and those words were not
thought to be possible to achieve in the
past they are possible I'm going to just
mention
two of my colleagues Tiana zernich and I
dindt around us we have a paper that's
now in the archive where you will see an
algorithm that does this it's a pretty
simple little economic based algorithm
where the decision makers are all given
a little bit of wealth and as they make
decisions they lose some wealth but when
they make certain kind of decisions they
gain back some wealth and it means you
can make decisions in a whole company
over a lifetime and I could stop you at
any moment and say how many errors have
you made up until now and that rate will
always be under control so this is a
brand new word it's part of me the
broader problem is she learning when
you're doing decision making ok that's
how it slides on that should I'm going
to skip because I've reached the end of
my talk that may move all the way to the
end in just a couple of slides here okay
if you don't control false discovery
rate you will make bad errors and this
kind of gives you an example of making
bad errors where you have type 1 type 2
error control really really good but
your number of false discoveries is
really really terrible ok so this is a
topic I hope everyone has learned about
where we'll learn about so here we go
party comments all right so let me just
pop back up to the right top this is the
error where everyone's focused on
pattern recognition it's being a
commodity you know tensor flow pipe
torques or whatever you will that's all
that's what they do it's pattern
recognition it's not decision making and
I hope you've seen a little perception
here of how important decision making is
for research going forward for this to
really be a field and for people to have
business plans which are not so quite
quite so broken thank you very much
[Applause]
浏览更多相关视频
Barbara Gallavotti | Che cosa pensa l'Intelligenza artificiale
Dev Jobs are being taken. But not by AI.
Nick Bostrom What happens when our computers get smarter than we are?
Algor-Ethics: Developing a Language for a Human-Centered AI | Padre Benanti | TEDxRoma
35. Che differenza c'è tra Intelligenza Artificiale, Machine Learning e Deep learning? #36
Top Skills for AI Era Success
5.0 / 5 (0 votes)