today is the day we talk
about is a company that started in
Cambridge and they’re pretty damn impressive they outperformed deep neural
networks on stage using a new system and today I want to talk about that system
and maybe a little bit about the team and the whole experience because Wow
literally the entire day was just filled with one PhD after another going on
stage and destroying my view of AI and it was amazing so hopefully I’ll be able
to share some of that experience with you guys okay so the company is called
Prowler this is the folder for the it was sort of like an AI summit think what
they do that was so revolutionary that’s this nice picture I wanna show you this
is the platform they invented this is it we’ll get into it okay so on that
platform they combined three elements that has not really been combined before
they combine probabilistic modeling reinforcement learning in a multi-agent
system that’s honestly it that that’s what gave them that leg up it was so
impressive he took I have to say his name cuz he was just he deserves the
credit you has ever you has SEPA is currently the head of operations and
product delivery and he took the stage in the end of the day to show what
they’d actually built the platform itself and ruku just blew our minds on
stage within like two minutes he had trained an entire algorithm to honestly
out from any deep neural network I’ve ever seen with two minutes of training
and all they’ve done before was set up the agent system they have to create
that and they said that it took them two weeks to
make but even if we take that into account it was so so they can see the
partners in the sense of not really doing anything useful about the bogies
that you may see at the edge of the sphere and then waiting to pick up the
fibers essentially we haven’t learned to play this game what that means is the game plays for us
then we are on the town together with the starcraft learning
process and then we have something that you’ve sort of seen already today the
taxi fleet bright areas is where the customers are appearing that actively
believe that they are in the city center so we have essentially a taxi fleet that
doesn’t understand the dynamic of the environment they don’t quite know what’s
going on the result is that the waiting times for the customers are longer
service is poorer at the top is matrix being reported by the hover over the
graph is going up which is the big thing will be more others that is receiving so
it’s finding different things that seems to be let’s make it seem to me that it’s
making great contrast here increasing with Ward’s which indicates that is very
better just to explain I’m cutting all of this together but he did all of this
on stage within five to six minutes he cannot have had the system’s training
for more than two to three minutes each it’s it’s insane they they were training
at the same time in the background on a cloud and we saw him set everything up
we saw him call back to it there was there was really no funny business going
on and it’s I just really need to emphasize how impressive this is they
train these systems two minutes on stage locations which obviously
to show the wait plans for the customers and better service for the customer so
so that’s exactly what we what we want this thing to do right so we do that so
we have an environment API which we have been developing to enable integration
with these environments so to lead the creation we need smart Roth game can’t
recall the exact kind but I think it was two three weeks okay so they know engaging actually
different frets and this is we’ve learned during my presentation we
started from the state pick up to this statement oh my gosh I should be doing they’re going to the dealing they had
the the standard leg test there you have a simulation of a leg and you have to
make it jump I’ll put it up here we call being a checkpoint which is just a
snapshot of the learning that we taking in the cloud systems while shifting goes
to the computer at the moment as soon as we have Elaine listen you’re going to be
launching the robot with that latest learn for policy and everything’s yeah
we’re hopping away into the distance happily and join the other robots at all so that’s all our free applications
weren’t very different decision-making problems during this presentation they
did that they made it like that it could run after two minutes of training that’s
insane like that makes no sense so we’re gonna
go too much into details about how this entire system actually worked and how
they create the agents first of all because they weren’t super specific
about that in any of the talks so don’t really want to butcher it and also I
don’t really feel like that is the interesting bit of it the interesting
bit is that they made neural networks way more data efficient which I never
thought would happen when they can train the network’s quicker they need less
data to train it which is the closest we can get to our way of making decisions
if you think about it you need to learn something you rarely need to make the
mistake more than once basic stuff like running into a wall
you’re not gonna do that more than once then you can realize that oh I need to
keep an eye out for these aspects and then you may walk into a glass window
one day and you realize okay I also has to keep an eye out for transparent
surfaces a computer would need to do that multiple times to learn that
current and it still does with this system but it showed us a graph on on a
logarithmic scale they were doing it in less than a thousand tries where current
neural networks are that needs millions of trials at least
hundred thousands of trials so they really managed to improve the learning
process immensely and that’s really important if you’re gonna make
decision-making which we can’t get a lot of data on a lot of issues especially
not the specific nuanced issues if you need a lot of the same nuances that
makes it really difficult to find data in most cases so this is the big do
really the big breakthrough with the multi-agent system they’ve put in and
the reinforcement learning because the the reinforcement learning isn’t really
a new thing but if you really want to learn more about this then I definitely
encourage you to go to the website this is the website they have a lot of
interesting stuff and a lot of papers if you want to read about the scientific
stuff I read them but I’m a geek and they’re interesting like beyond belief
so I definitely recommend them but just be prepared if it is very advanced
mathematics so yeah enjoy and thank you very much for watching this video if you
enjoyed it and you know someone who might enjoy it too then please share it
it really helps me out and yeah I’ll see you guys next week
take care yes we shall from prowler dot IO so the full name of
the companies I keep crawler dot IO and or problem because if you guys do a
search you’ll find out boy actually before I saw it so we are the world’s
first principal AI decision-making platform we use machine learning to
understand guide and optimize millions of micro decisions in complex systems
and specifically complex and dynamic systems like massively multiplayer
online games smart theory simulations and the key word here is principled
because what we are doing here is using interpretable principles of
mathematics for learning and decision-making from AI perception is a
solved problem computers are much better than humans in identifying that a dog is
a dog and a portal is not a tree the next big challenge in AI this putana
most decision-making Colorado iron the world’s first platform for autonomous
decision-making with our platform and our AI BOTS we are able to make
autonomous decisions using reinforcement learning compete and collaborate with
other BOTS using multi agent frameworks and learn to mimic humans using inverse
enforcement learning as of today all decision-making is done using
handcrafted rules and a problem we believe what can be learned should not
be hanged browser

, , , , , , , , , , , , , , , , , ,

Post navigation

8 thoughts on “MACHINES Make BETTER DECISIONS Than HUMANS | AI Summit

  1. Dude are you serious? The audio was bad before you drowned it out with lame music. Great content. I wish I could hear it.

  2. It's been a serious cliff-hanger since your last episode! What is the name of that company…. – thank you! Need to spend some time digging into their site. Keep up the good work!

Leave a Reply

Your email address will not be published. Required fields are marked *