Transcripción
Extracto de la transcripción automática del vídeo realizada por YouTube.
hey guys good morning yeah so I will talk about machine learning the promise and all that was just to get the cfp expected so I accept it so don't expect too much away so i started with machine learning about six months to a year back and this world like
amazes me with how much stuff you can do and how awesome it is so at the end of this talk i just wanted to get excited about that too and go and try it out you can read more about me there 21 talk about that the slides and the code that i'll show are up
on github and slides and all that yeah so i was preparing this thing and then they just yesterday everybody knows pages here right he had couple of jokes that were exactly what I had and were like oh god I have to prepare new jokes now but at night I couldn't
so it was too late anyway so I didn't so my jokes are still lame but still laugh like you did for his talk right first of all happy new year has anybody ever had 200 people shout happy new year to you at the same time no want to try that all right so on
a count of three right one two three happy view yeah that was actually much better than what I expected because I had two more steps after this I'll just do one of them anyway let's try even better right let the hotel know that we're doing something
awesome in here so shout happy new year and stomp your feet don't worry the floor is pretty strong except probably this part so all right like this three times right so one two three happy view yeah all right you guys are already awake I know but anyway
so we'll take a very brief look at machine learning em there are many awesome talks that tell you what machine learning can do this is not what this talk will do I'll specifically focus on classification which is one type of machine learning and an
algorithm that you can use to do that which is SVM and then at the end we will see how you can do this in Ruby so yeah this my first talk and I thought I should do a live demo right before we actually go so what we are trying to do today is we have all these
large data set of about sixty thousand images that look like that six and five and seven each one is 28 x 28 pixels and what we have is raw grayscale intensity so you have 0 to 255 and there is this data set available that you can use to practice your algorithm
on or train your algorithm on so it's hard to read it's all binary and I don't understand binary so I put it out in CSV like this and as you can see most of the pixels here at the top bottom sides they are all going to be zero and in the middle
you'll see some numbers that's going to range between 0 and 255 so what we're trying to do is take these 60 thousand images basically create have a algorithm learn those images and then when we give it a new input like another image that sort of
looks like one of these or maybe a little dissimilar it should be able to tell us what it is right so that the MN is tea set also comes with 10,000 things for testing so first we'll run it through that and see what our accuracy is and then i will show
you some of my bad bad handwriting all right so i will talk about most of these things later but i am basically you don't need to know what linear is i'm just showing that i'm doing it on sixty thousand samples and i won't do the all 10,000
right now i'll just do 500 of them and that's going to give come back and say hey I'm eighty five percent accuracy and I didn't train the algorithm right now because that takes a lot of time I already have it saved and you can see the code
and all that the saved models and github but for now the main important parts are this algorithm it says is eighty-five percent accurate on the 500 samples that we tested and all these things that you see like for example this guy we got two images that are
actually nine but I reported as zero alright so those are bad everything that you see along the diagonal are good everything else is bad and I didn't tune the algorithm at all because I want to keep it simplistic and all that but you could do a lot more
[ ... ]
Nota: se han omitido las otras 1.967 palabras de la transcripción completa para cumplir con las normas de «uso razonable» de YouTube.