Embed code
Note: this content has been automatically generated.
oh well mine i mean i work in the
c. d. is elaine omission centre imagery spain
yeah today oh i'm going to compliment ah beside us talk about it and so slow
and then we're going to to talk about three taste
also use a using this machine learning library
i first i will ah do you ah classification problem
comparing a psychic learn code into the tense are
after that we will these into the analysis based apps challenge
uh focus on that ah i won't these a
year uh doing a program with a for classification of an impact pretty soon for as there right
and after that i will be eating she'll uh the mixed system reverse engineer brain
a project that i'm i'm working on it so how many of these are
i'll actually working with some machines are there any library river has
okay nice so um former former these uh interaction
i'm just going to want to talk about the circle come pray tour
to see how easy ease to make like any kind of
classification problem you since i keep uh out there
and the a. p. i. often surfer for that i will use the titanic data
set trying to classify how you're probably t. to downright or to see inc
depending on the data of the passengers okay so i misinterpreted
file like a cheese is all the data does survive
h. and class uh in that data set why because ah
yeah with the titanic has signed a people will say like okay lest oh let's uh
safe first the tile and women are man and after they well i'll
look in after a at the fair or the amount of money
that the person you're we're paying for for the dated panic uh i'm bart said
something that i have a first i will like eh
comparing call for the uh second there and use it and that's for for importing like
the data and then i'm going to use like the one in and or like it's the regression model
and then it got also going to learn all like the matrix
that's like to learn kind deep gave us about the classification
uh i'm uh putting like the data uh as we teen
uh the training and the test they date that because this is very important in machine learning
we have like a data set and then we test data and then
we say o. k. a. from the side you're going to lair
from the training data and the test a you're going to test
if the state that uh ease oh it is good
i know or is not so at the end i owe us is split it with aside in a random state
but yeah the are the like these the logistic regression with our
so i could learn is no longer than one one line of
code okay like taking like this feat for the trainee
and we print accuracy how well is more of classify and he's doing as you can see
here we have like that all point sixty seven this is the first approach of
so i did learn as me file at all in the in the previous talk if you remember that it's an classifier
we have like an all point forty nine occurrences score
so are we should like also like be some kind of exactly yeah um compare how
well new libraries of machine carotene are doing in compare with tens of role
with tens of harm what i'm going to do you i'm going to import but they're here
but i'm also to import the web metrics because i'm in this approach
uh tons of low is a very like photographs of me highlight our cans eh said and this
room in this very meant i wanted to to put like all the accuracy to test the
a so here as you can see we also like test uh is breathing today training and test data
but here ah we use of the learn the linear classifier
that i haven't had a heart also like a lot done before and we do that
up to my server in the up there uh we use like the learning rate
but it means the how many are times we are going to you tell the
networks you learn from something okay so i'm here how we of freely
and yeah yeah we change the l. s. text the training
is that and they bought size that means that
i'm going to u. r. and put the data or far trainee
and eh tell like as the house instead okay there from a is there from the is there from these ideas there
is like oh i'm computation me doing this that and the learning rate that gave us like the the accuracy
so on this first experiment that conclusion was that okay
oh for the love first approach secular device like more accuracy
but when i was saying will to relieve feed their tents or flow model it works better
oh no it's you talk about our detector how important is to
seize the data and to get the data i mean
uh oh the poor is not um my dick box in which
you can pull on your column surrounding tissue are we soul
they think that has to have be or there and has to have i mean i'm in a
few months ago necessary east oh their main core
problems the they are uh trying to solve
once in here and make that like open for you know a tours all over the wall
one of the main our topics they cannot was to you oh
mm prevents make a program that could uh pretty eat
and classify of me was near or with bill objects
around the around the solar system a cycle
and not a nice they can be like a problem uh this classification
problem this problem is very see a human they have to be
like of serb in all the time they asteroids or the meals
so i'm going to tell you more or less how what
they at first are a no no of their battery or whatever and they all
see something okay they may i'll i'll classify a
base owned several points uh they see
like double c. t. the call or upset there and then they wait out
for awhile and then they officer that as their way through the time
so ah that problem here was how can we do something machine learning
how can we train them i thought that really light gave us like out all these uh human uh work
so we're well uh this project that was cool a beep us the right that
maurice though was that was an hour to get to our uh neural network
that could like help us to to classify i uh
we yeah uh the other night oh six technology geeks on that we
are basically are we're passionate about the potential from our machine learning
and we believe deeply uh in the power of producing knowledge to shoot data
how we knew all we have like the uh this problem of misery in the arts
and we wanted to pretty beat but we also wanted to classify how these uh could work
so at the end uh we went to openness arc that easily hewitt a up a
a platform that got a lot of they don't would you can like vibrate
and it's like the like the most like significant data sets
but who happens to you deal with this problem
uh one of those was the potential hostels as though it's a that that's it
that means like they have like an only me all the where like less than
all point all five has only got a synagogue units because that's the
the order that they asked to really has to our impact on earth once it's it is a those near
and also like the the um the more it's there is like the mini mm armies intersection
distance that we could have like with uh the errors and eh yeah and then you
uh we yeah as you know at the end like a a a and there really is no
no more than an input and an output and we use like several kind of functions
she'll i'll try to she is and give 'em the that a now don't i mean
so what we you know this is okay let saddam less used like o. r. three hundred
take 'em function and wave functions because that's the kind of uh oh our wheat
or trajectory that an asteroid cousin half and the l. e. r. t. team though
out that there were and how how the past it was was movie
so i'll sit me try that has told uh there is actually a surety
and here when when you're doing a a neural net and we were like oh i studied a lot
with mathematicians uh and uh the experts of the feel hot was a
these going to work so i the first uh late here
we were eh with these putting sell hasn't as though it's a
they thought and that is they came up like i classification
and that was like the happy and apollo astronauts those are the ones
uh that we want to you to this study because our there what entirely a dangerous to
impart oh and your bus but as i told
you before um a astronaut meek of
i thought i might feel and astronomical eh that's what we are
a workout with time review if you a way to die
and then you classify again there's the right you might change so
what's important for the for the neural net to know
then we were created another a year it was a a talking about how that time could impact
in the in the us that classification at the end we also like to use that i thought it
was so for you to know it starts with physics okay this is the physics are beta
then we have like they classification with a human knowledge and then we have like they get more chemist
a kind of classification because also when they're asked to come here by
you can see more things about it like the core or so um this is
like a a parity for classification that we put into the into the inordinate
of the uh be okay like tons of low will ones do
they say to you that your neural net is easy
and they have like put the uh based straightforward part that's that's okay but uh you'll have to you
a also have like some kind of a craft might see if a process
uh so i could leave yeah if not where you say okay i want
to sign an hour nap about whatever it ends as their rights
dots wherever at first i was that yeah that a really like a study that
that's it i'm fine a data set that goal from the general knowledge
to all the most out great or small knowledge if i want to classify
cats for example of details at the first layer i'd like eh
the eh uh pictures of uh get there that day and the for lex okay
and then go wow and then maybe i'd only the phase of upkeep and
and then i go up again and then it would only like their hair of updated so if
i saw a picture of a key then there's negative on depends on the cage okay
as i say i have been training all on your own a it didn't really recognise it
this is very important because if you don't know you know that in
only train like image of okay then we for our uh
without pay them for letting you get on our face of the gate and say this is okay and i don't see like
a tail and see so this important to seize the data to be
the yeah i mean it when you're training a neural net
you also have a lot of protocols a in this case we have like when you're recognising protocol that was already apple
that uh i i was too prone a lot of the car and be aware
of uh taking off a lot of the that this is a also important serene
and urinate because if not you are like a put a lot of noise
okay as in these are here to yeah estate off or or training
heart beat at the end this key because it not you do you could be doing all long
i'll be aware of that the uh the light years have to be able eighteen
each time you make a liar okay yes uh probably a in fact tons of low uh it's very
good for that good you can slice it insert or one of the main core units units
and then i put light second there for like a
year whenever you my whatever you want i want
but yeah i'll i'm uh we were we were already and uh one room between
now i'm on both twenty five a best projects are on the wall
in the first uh five for best use of data
right yeah mike aren't as a cool i say okay guys is great because you are like
really really really gain knowledge into a into an it not just like
a seasonal model and probably you i'm giving all the data
and i'm going to talk about what i believe is there are the future of of tons of low
it's more like the uh reverse uh in your uh part of
uh for that i will upload some on uh or that was a real story
that happened to me and how and how and how i deal with that
so um have you it's a goal i happier a troll on the internet in my two year how
it was like with this massive for me and i really really really hard time
and i i i i start to think about the warning us of they turn it on what is that digital wall
i'm in pro and over the human emotions
but i also was like getting to meet and i
get into the our goals busters batteries a case
in which she was like as much into here and some of the ah i realised that it was not i'm
and that uh we were like of implementing hate
a with b. data with data over the internet and what we really are doing without
so i was thinking about how can i i'm the do absolutely sure that really like help
me to to deal with that using maybe like okay yeah uh seventy to eighty channels
but also like artificial intelligence to to transform that we yeah
it to you and to to create something something new
that starts you to you all work with a with a number of
that uh it could help me like to identify those tools
z. you what was happening in like uh on the internet we
with eight and not i um use something uh about it
and then i i realise that the whole i'd say is that they thought was
very important so i'd like to meet basic all five emotion classification okay
uh the last phase like dealing with the during two thousand years and i'm
just uh taking like well what i believe there are high quality models
and implement the implement them into programming language okay so when you talk
about paid eight a. is an emotion that doesn't come along
as far as a hate at times uh uh in a sense of abhorrence of
a boring as a state and it can be like even transform into rage
uh so i created like oh with uh wait a back door thoughts of words
but i believe was like these um a plaster of uh how these emotion
was uh was working from the uh low level okay that was
like those other runs to the high level eh rates
and then we get like darby star tense our own known tons of not
uh we felt like the that much tomb gradient of the motion
oh reckon saying of these okay fine see like oh input other right
a that means that that that person i'm waiting he's oh okay or awards onto your
and then i i kept waiting to there there aren't there hate or there right
ah here is the mathematical model me try that told you
about um that the seas there for the classification
uh as we uh as he told you well uh we were
creating the brambles uh with old ways and i'm bias
and uh we will also like i would be that it's really may
show ten sample how that like the the points of information
is that like oh all these when you're doing a classification of our model and this is what oh out with me
try w. well about the this a multiplication between our the
by asked um we we soon like the bean
what a number are complex and things are not linear
so that yeah like the first like eh a grail of complex c. d.
that you can like give to enter network is making like convolutional
okay so have like these other more let me try that has that told you about
about the this linear function that draws the light you can like apply sigmoid function
but what does is like it transforms the light into
an exponential near well then it a pattern
so from this yes only with these are like well you
transform i it into an arrow um eh classification
also a place holder and stuff like there is a hard to take this space there um do you
need a break even on the court cross entropy like oh me try that they've you talk
about this so much but i think this is like a very important because it tells you like how well
the neural nets do you how well these uh and larry
and the some of the medical function that uh has like these um
the these uh aim to be at the end uh it's very important as
you say it like between instead and how we hope to my seat
we'll start decision because as you know like we're always like competing
wraps and we collect the day pretty sure we've oh on
the evaluation has two steps that are like quite important we correct yeah
then around and we see like that we see that it has
that's why i don't believe like deep learning is a mighty black box which you
put everything and you get like a resorts you have to be light cease
you are at first i will say like okay take like are they tell if
you know and we we sold you might know it as well okay
so you're he yeah and you correct upper these humans and they accuracy that u. k.
so then you can escape employed a lot of data so making like a small first and then a scale up
but you know no sean is not all the that is not like that state
that can like it's not picture motions are things are moving
on and and uh have like their their complexity
and i was like thinking how could i oh i'll start the that
the motion better and try to really really understand the troll
and now on only like make the picture of the troll in a separate time and try to to understand
better to saint uh in um more at great
and um better way so um i was thinking you know a lot about uh bats
because uh our division to inhabit a really out of maps ain't on it you know
what i was thinking about the um they eh imaginary numbers my numbers are great
because there are like that's that things that that's the next east
but when you buy it together they provide we'll things
or is scott r. a. s. colour eh but who's so
idea like i'm uh i was thinking like emotions
also i work as the majority numbers because we have
like they are emotions are like around as
and at the end like oh you drive us to an action that we can measure
so i'll i'll also i was thinking that emotional a hustler gradient
itself before you've got any movie showed you kind of
feel something you have like there but but i it's for you forget it
not like some kind of sensation so at the at this trial you could see okay i have this
notion but before even in motion or before being strong enough for me to to say something
oh it might be some kind of sensation that i could not even on the internet
so yeah i could i could do an awful in which i
got a see some tools actually are not swells yeah
so i can we like these kind of tense or a having like big
great of our of of on motion but also play their motion itself
that you know that like these matrix in which i can put like
how this emotion like uh wasn't working at every complex level
yeah i'd yeah it was great because i would like to do about
numbers like that to get to that sounds unlike loopy telling
to the okay guy i like the racial you cannot be like that and they might you know number like okay like
mm get real don't be like that so id and it's like
a a it's on um i max a conversation in between
in between numbers that you can has to say to you
to compare it to to robert's too complex uh problems
so now i'm okay the it gets better because i can like uh would like these neighbouring e. ah also
like i can like make these uh temple all in with i can put a complex number x.
and i can read also my controls a date i and i can like start
like making like bills downstairs again that have like that a complex eh apart
eh uh this uh complex part and i'll get you i don't know
if you have like been working out of we fed that's awful
right yeah and uh one of the things that i lost all this is that you can like
a size also the type of tense or that you can have
so for example if you're working with images you are going to
use hind i. t. thirty two in this case as i
was working with a complex number with this colour i was using like this kind of our uh this kind of things
so um what we were doing was like we were created a neural net
that was i given as surveys okay so this is i think this
they are the core future or or forty visiting tightly and now
and you'll hours programmers can create were like a like a neural net as a service
that go like be on this uh whereas the service and you can
like sort of a problem um uh you seeing using base
oh i don't know it's me kinda has a told you about this but if you i knew
he uh to tens of uh oh oh good like to you gain too much you're again
i recommend you like these uh oh that's awful later dot o. r. g.
the tissue like baby usually i'll uh with these other in it
and see the lay years that later suffocated we're protein
the relationship out of the data and the kind of problem that you want to solve it easy like out approach
to how the problem is going to to be heaven with is very useful ah and i
always uh when i think about the number that you know it for starting to call
ago here in se more or less the classification and see what davies has like a meeting
or a mom it's coherent so if you guys i neglected going to into neural nets
i really like a recall main this uh this uh website
and oh here i put like oh well it's a male than it in terms of goal
because okay uh are you you have see like arose data but how can i call
a neural net ah id fly the and run it and then ah
the most important thing that uh you're going to be dealing with and think you are the a heathen units
which may guy that has told you about also active in your own function we have seen
the asteroid exercise are we're dealing with that whole b. c. seemed like the eh
they they functions and re creations that we're going to have like kind
of the final classification problems the big they are going to solve
also that things like the the rainy day larry make the what the end
okay you have like the target audience village does that problem and the
layers of a cake and the function that i'm going to half the the
do you think that i'm going to put and the learning rate
and then i started thinking again that he is and say okay so uh
emotions are complex eh more complex even than that
because we're we're like hitting the internet
either it is a are flowing data all the time and is
changing and artificial intelligence um should be more light up each
are also problem and not d. n. like a brain is oh
um it's a a neural net of the neural nets
what i mean is that we connect a neural nets and we create a brains
okay so i d. and i think this is the problem
that artificial intelligence uh wants to really realise all okay
so i said let's say anything with my i'm actually i'm hitting with my my here troll
um i was thinking that at the end i wanted to have like a negative emotion
and i didn't want like a qwerty level of tool in or wherever i went and fear are real
motion on the on the stroll i want to change i want you to give all feedback on
that's for that a minute if uh to meet or met if
bird there is common into ah into a positive uh one
so ah limits is there is a part of the brain
there is like be on the new cortex okay
eh that basically uh it deals with our emotional selves
oh then you're cortex eh i mean dries for general intelligence worries
like oh oh memory and processing but linguistics and is different
it deals with uh oh we're strain for or couple eighteen
danger with uh dealing with o. o. where emotions
so at the end is a part of the brain that uh is like processing uh processing that
ah so ah ah i started seeing how and small oh and
you wear off of how how brains uh really work
and how these part of the brain at the end is like a influencing daniel cortex
and driving us into that ever a motion eh that's that's
generalisation and at the end we have like a structure
that would uh we could like impair or try to do our reverse engineer of that
so are the and uh and uh it's the it has like four main core parts gate
and and different a kind of a impulse we have that animals
and sensory cortex what isn't a recorded we had only for the front part and i mean that
and then we have to kind of different inputs that comes here
okay what we call sensory inputs i mean you might chase
we deal out of images you know where emotional brain wellstone emotional sickness
singers that i interpret their line my mathematical here models
so okay i i it's it's great yeah because i take
like the intention already like want training around it
all school that has without tones of images and i like it
to my emotional mathematical model created more complex a data sets
q. to deal with that uh uh for the it didn't say a um okay so
with input i have like images i have an emotional signal and then have
a a box that you know guys that you is not longer box
and then i have a uh let an output that is going to be another tweaked
this to me is how a is going to have on emotional gradient that goes from the being assertive okay you're
trying to me but i'm going to be as i'm going to put you on a circuit a to e.
jill compassion because at the end like a haters ah
or her like um are like oh own move
i'm so lost souls that have something to say and we shouldn't like just keep eight
apart we should be soon and we should uh also like so some compassion
about bass and now i'm dealing with uh then it's a the next
time it's for the is that is getting they they they tap
yeah i really need because uh i have like the insane shown a model data sets
in which i have like features of of trains or t.
tens or dogs right need a pictures of joy
i need pictures of being sat i need pictures of hate
so the data set that i mean it's like why mm specific for this
and there has to be huge so oh mm i'm a
i'm a working every day all on building this
eh but i yeah i also need like a tones of that uh burbank kind of a beta
that is giving me like those emotional emotional signals so um i
hope to one day like finish my my emotional brain
uh in the mean time if you will have the opportunity to see all of we get
with a troll a side eat uh please um forgive
uh but also stay calm okay and if eh it with
a heart she'll stay just away from from the phone

Share this talk: 

Conference program

Jean-Baptiste Clion, Coordinator DevFest Switzerland
26 Nov. 2016 · 9:40 a.m.
How to convince organization to adopt a new technology
Daria Mühlethaler, Swisscom / Zürich, Switzerland
26 Nov. 2016 · 10:14 a.m.
Q&A - How to convince organization to adopt a new technology
Daria Mühlethaler, Swisscom / Zürich, Switzerland
26 Nov. 2016 · 10:38 a.m.
Animations for a better user experience
Lorica Claesson, Nordic Usability / Zürich, Switzerland
26 Nov. 2016 · 11:01 a.m.
Q&A - Animations for a better user experience
Lorica Claesson, Nordic Usability / Zürich, Switzerland
26 Nov. 2016 · 11:27 a.m.
Artificial Intelligence at Swisscom
Andreea Hossmann, Swisscom / Bern, Switzerland
26 Nov. 2016 · 1:01 p.m.
Q&A - Artificial Intelligence at Swisscom
Andreea Hossmann, Swisscom / Bern, Switzerland
26 Nov. 2016 · 1:29 p.m.
An introduction to TensorFlow
Mihaela Rosca, Google / London, England
26 Nov. 2016 · 2:01 p.m.
Q&A - An introduction to TensorFlow
Mihaela Rosca, Google
26 Nov. 2016 · 2:35 p.m.
Limbic system using Tensorflow
Gema Parreño Piqueras, Tetuan Valley / Madrid, Spain
26 Nov. 2016 · 3:31 p.m.
Q&A - Limbic system using Tensorflow
Gema Parreño Piqueras, Tetuan Valley / Madrid, Spain
26 Nov. 2016 · 4:04 p.m.
How Docker revolutionized the IT landscape
Vadim Bauer, 8gears AG / Zürich, Switzerland
26 Nov. 2016 · 4:32 p.m.
Closing Remarks
Jacques Supcik, Professeur, Filière Télécommunications, Institut iSIS, HEFr
26 Nov. 2016 · 5:11 p.m.
Rosie: clean use case framework
Jorge Barroso, Karumi / Madrid, Spain
27 Nov. 2016 · 10:05 a.m.
Q&A - Rosie: clean use case framework
Jorge Barroso, Karumi / Madrid, Spain
27 Nov. 2016 · 10:39 a.m.
The Firebase tier for your app
Matteo Bonifazi, Technogym / Cesena, Italy
27 Nov. 2016 · 10:49 a.m.
Q&A - The Firebase tier for your app
Matteo Bonifazi, Technogym / Cesena, Italy
27 Nov. 2016 · 11:32 a.m.
Hasan Hosgel, ImmobilienScout24 / Berlin, Germany
27 Nov. 2016 · 11:45 a.m.
Q&A - PERFMATTERS for Android
Hasan Hosgel, ImmobilienScout24 / Berlin, Germany
27 Nov. 2016 · 12:22 p.m.
Managing your online presence on Google Search
John Mueller, Google / Zürich, Switzerland
27 Nov. 2016 · 1:29 p.m.
Q&A - Managing your online presence on Google Search
John Mueller, Google / Zürich, Switzerland
27 Nov. 2016 · 2:02 p.m.
Design for Conversation
Henrik Vendelbo, The Digital Gap / Zurich, Switzerland
27 Nov. 2016 · 2:30 p.m.
Q&A - Design for Conversation
Henrik Vendelbo, The Digital Gap / Zurich, Switzerland
27 Nov. 2016 · 3:09 p.m.
Firebase with Angular 2 - the perfect match
Christoffer Noring, OVO Energy / London, England
27 Nov. 2016 · 4:05 p.m.
Q&A - Firebase with Angular 2 - the perfect match
Christoffer Noring, OVO Energy / London, England
27 Nov. 2016 · 4:33 p.m.
Wanna more fire? - Let's try polymerfire!
Sofiya Huts, JustAnswer / Lviv, Ukraine
27 Nov. 2016 · 5 p.m.
Q&A - Wanna more fire? - Let's try polymerfire!
Sofiya Huts, JustAnswer / Lviv, Ukraine
27 Nov. 2016 · 5:38 p.m.
Closing Remarks
27 Nov. 2016 · 5:44 p.m.

Recommended talks

Torch 2
Soumith Chintala, Facebook
5 July 2016 · 11:21 a.m.
ACM-W Athena Lecture: Large-Scale Behavioral Data: Potential and Pitfalls
Susan Dumais, Distinguished Scientist, Microsoft and Deputy Managing Director, Microsoft Research Lab
23 April 2015 · 8:36 a.m.