Learning to See [Part 10: World Domination]
Вставка
- Опубліковано 19 вер 2024
- In this series, we'll explore the complex landscape of machine learning and artificial intelligence through one example from the field of computer vision: using a decision tree to count the number of fingers in an image. It's gonna be crazy.
Supporting Code: github.com/ste...
welchlabs.com
@welchlabs
![Learning to See [Part 11: Haystacks on Haystacks]](http://i.ytimg.com/vi/biy2yU3Auc4/mqdefault.jpg)
![Learning to See [Part 11: Haystacks on Haystacks]](/img/tr.png)
![The most beautiful equation in math, explained visually [Euler’s Formula]](http://i.ytimg.com/vi/f8CXG7dS-D0/mqdefault.jpg)
This is some of the best content on the internet, if only it was a team making the videos instead of one person! We need more of this stuff
Álvaro L. Martínez totally agree
This channel is so good. It feels wrong not paying for something this well made!
The coffee part was so satisfying.
It is amazing to think that you are able to do this videos alone once every 2 weeks.
Thank you very much for this amazing content!
Thanks for watching!
Will it generalize? :D
6infinity8 Because it is not specific to the data set, please watch the whole series
I follow the series since its begining, I was just laughing at the crowd's voice he added ^^.
6infinity8 Oh, sorry, I missread Why will it generalize?XD
sorry
It scared the living daylights out of me lol
Yeah, this makes me shocked. 3:40
I just saw the thumbnail and my primal insticts told me to click as fast as I could
Hey Welch Labs, what is the outro music? Really love the groove!
Professor I have become such a big fan of your videos I can not describe. Great content and such profound knowledge!! Thank you for this. Currently attending the CV class at UNCC under you.
wow! for such a specialised topic you're visual editing is on point! thank you.
Thank you!
Studying for my Machine Learning test, although it's a great topic and really interesting I'm ready for the holidays ... the "WILL IT GENERALIZE?" cracked me up. Needed this, thanks :)
This is probably the best put together series I have ever seen on UA-cam. The quality is so high, but the videos don't take forever to come out. Keep it up!
Thanks Adam!
Noooo!!! I have to wait? Just watched all videos on this channel and didn't realize that he was not done with the current series! This channel is soooo good! Its like waiting for the next book in a series. Seriously, he has really covered three topics that I have wanted to know more about. thank you so much and I know that the videos coming in the future will be well worth the wait!
Working on parts 11 and 12 now! Thanks for watching!
Welch labs is the best youtube channel on the planet.Amazing talent and creativity. (Read in Trump voice)
Hahah, thank you!
I was running off of a batteryless, screenless, keyboardless, laptop with a broken trackpad that couldn't even boot windows because the integrated graphics chip had gotten fried.
Then I fell in love with machine learning, and after buying my rig I came to realize I had learned to love myself as well. I did buy a Gaming PC, bro. Best decision ever.
The series is awesome. I'm happy I've ran into that. I'm little lost so far but will handle it eventually! Thx!
One of my favorite channels. This series is as much of an introduction to Machine Learning as it is one to Computer Vision.
it wouldn't be a Welch Labs video if it didn't end with a cliffhanger ;)
This channel is awesome
You are really good at this. Keep on producing such high quality content !
Thank you and thanks for watching!
this series is getting better and better!
bro , you Lionel Messi of Machine learning!! You are the G.O.A.T! I have been struggling understanding these fundamental concepts for almost a year! But you just....I AM JUST VERY EMOTIONAL RIGHT NOW, I JUST WATChED TILL THIS EPISODE IN ONE SINGLE SESSION AND I AM GONNA FINISH IT TONOGHT
Haha hype it
QUick! We have to conquer the world with machine learning before anyone else watches this series!
Another great video, I took an excellent a bit more mathematical course at my uni in Cambridge, but this explains a bigger picture much better, since we just jumped into Neural Networks straight away. I am looking forward for next video and more importantly code for the videos already in this playlist. I learn the best when I code with you at the same time. Keep up the good work.
Thanks for watching - I'm working on getting the code out soon!
The quality of your content is ridiculously high, thank you :)
That's some good coffee : )
dang it, I was just starting to feel like it was time for another one of your videos and was getting excited. now I have to wait AGAIN for the next one
Great video! What music did you use in the outro?
This channel needs much more views
Thank you for making these videos. they're amazing :)
You are an amazing UA-camr, I'm subscribed to you for quite some time and your videos are both funny, interesting, informative and well made.
You deserve much more subs and views!
Always pleasant to watch your videos, keep going !
Thanks!
"You are a bigger tease than a"
yay! finally generalizing!!!!
keep up the good work!!!
I realy like your videos! but I'd propose a small change: can you make the bottom right corner funny sentences more readable? the youtube buttons are over them :)
There are 81 choose 4 ways to choose the pixels and 2^4 ways to check if they are true or false. By that way there are 26619840 ways to choose a rule. Since there are also about 5000 squares you have to check for each rule, the final number of operations the computer has to do is about 133 099 200 000 operations, which is a lot of operations. Especially for ugly code written in a slow language like python.
It is using numpy though, so perhaps it is not dramatically slowed down by Python's sluggishness.
Now I've had to make a coffee...
and it's 6 pm...
so I won't sleep till 6 am...
really thanks alot!
Sorry?
Can you put your little sidenotes a bit upwards? So one can pause the video and read them without the bar in the way.
Hell yea another exciting video from Welch Labs! We'll build a robot to dominate the world which identifies coffee and drinks all of it!
Amazing series!!! Looking forward to part 11!!!!!
The 3-pixel based algorithm is like that , with its 2 edge-detecting pixels at different heights, because you showed it right hands images, with fingers slightly sloped NNE(up-up-right) at an average of about 15°.
Why would you ever do succinct series? This one is awesome!
I died at the coffee and "Bro, maybe you should get a gaming PC?"
:)
Why not use a CNN - aka: a Convolutional Neural Network with certain combinations of kernels that stack up to fully connected layers and then train the whole thing with Backtracing Algorithms, or is this what this series is going for? If so this would be the first time I actually got someone to explain to me every aspect of it and the reasoning and necessary steps behind it.
So keep up the good work and hurry up please, because I never felt so excited to get to learn more about machine learning.
notification squad!!
Your vids are class man ! keep on :D
Generalize means Empherically if the accuracy and precision and recall is a x percentage on train data , the same metric should be on test data also
Bais is when we are puttin our faith on selected pixel value and trying to get the rule selected.. variance is when we are making more varied rules..however the desied to moved from bia towards variance and find the optimum is motivated by the fact of emperical evidence that when we try to get the prceision and varaince from one pixel rule to 2 pixel rules or more.the confusion matrix's metrrics give u more postitve resutls..and hence we try to strike the equilibrium. Also may be the metrics of 65 percentag is sought after @5:14 when we try becasue a normal distribution 68% of data lies near mean(i.e we are achieveing aorund mean)
Coffee outtro for the win. Awesome kettle, BTW. :P
Thanks!
I think your lessons are brilliant!
If only content was presented in such a clear and intuitive way when I was in college/school!
Is there anything I can do to help your channel grow?
Thanks for watching and for the kind words! Stay tuned for a Patreon campaign later this year.
amazing series! i love it so far, thank you for spreading knowledge :D
aesthetically pleasing af
0:50 no! a=0 and b=1 is same as a=1 and b=0 is you switch the X's
5:58 same, you only need the amount of 1's! no need to check each binary combination like 0101 or 1011
Wonderful video, same as always!
You're my favourite human right now :)
This may be covered in future episodes, but... if the machine is "learning", why do we go back to scanning through all the rules all the time? If we assume the previous 3-pixel solution is a good baseline, we only need to test the rules that contain those 3 pixels and one of the other 78. Surely that would be much quicker?
second video i ever liked on youtube :)
Woohoo!
I'm waiting for that *efing* succint series!
What the hell bro. I lost u like, a septillion videos ago and its still this series. Where the hell r u going with this?!?!
Outro music/song:
www.premiumbeat.com/royalty_free_music/songs/what-s-the-matter
Thank you. For some reason he refused to tell anyone what it was, so you're the real hero here.
Where is the music in the end from? it sounds beautiful!
Maybe I should try making my coffee the way you make your coffee, by pouring the beans into a filter, then pouring hot water over them instead of running it through my coffee maker. Or, since that's pretty much what the coffee maker does, never mind LOL
Oh god, I can't wait 2 more weeks! 😩
That's a lot of work for one cup of coffee, but I'm sure, it taste awesome
Now I really want to see the training set, and scribble this in PyOpenCL...and I hate Python. But it does look much nicer on screen than my C usually does.
There are 2^4*(81C4)=26,619,840 possible rules with 4 pixels, 39 times as many as with 3 pixels.
Wow, a data and coffee lover! nice! why didn't you time the brew?
"A lot of you are really freaking smart." Well there's the people that watch this and understand it, and then there's the rest, the guys and girls that watch in hope of getting something to know, for once.
it's so slow because it's exponential right ? Bc the 1 pixel one was to the power of 1 and the 2 pixels one was to the power of two (squared) etc...
I know, why it slow. There are 26,619,840 possible 4-pixel rules. That's 39 times more than 3-pixel rules, and over 2000 times more than 2-pixel rules.
If the fundamental problem was that it was 39 times faster, the comment that "the 1, 2, and 3 pixel rules didn't tank anywhere near this long" assuming "this long" referring to 37 seconds would have given an upper limit of 37*39=1443 seconds with an average of 14.43 seconds per 1%, assuming the algorithm was efficient.
I fucking love this guy
waiting for next video!
Fantastic video as always! Could anyone tell me what the song at the end of the video is?
You're awesome!
why don't you write this:
for i2 in range(2):
for j2 in range(2):
for k2 in range(2):
for l2 in range(2):
yHat = np.logical_and(np.logical_and(np.logical_and(trainX[:,i] == i2,trainX[:,j]==j2),trainX[:,k]==k2),trainX[:,l]==l2)
#btw,i really like the imaginary numbers are real series
Good question! Using nested for loops actually results in searching redundant rules - using itertools allows us to avoid these.
You have one of the nicest hot water kettles I have ever seen.
Why thank you. It was a wedding gift.
1:15 why only 1 winner?
There could be more like a right edge detector.
wow I didn't know machine learning had my jam in it
5:43 I'm like "well that's not too bad. like 70 minutes at most"
I have had way worse runtimes. e.g. creating a 3000x3000 numberphile sandpile pixel image. or reading bytes one by one from the hard drive (I fixed it very soon).
Fantastic! What was the song at the end?
you should parallelise your algorithm, would make it go faster
It won't solve the fundamental problem with the algorithm's time complexity though.
asdf that completely depends on whether or not a non-bruteforce solution can be found to the problem
Maybe the brute force part can be optimized somehow. All i was saying was that 1v4 cores won't make much of a difference when switching to the next level of variance. Besides, if code optimization was a key issue, why not write the entire thing in c++?
asdf execution speed is just going to be a polynomial factor, in the grand scheme of things, a fast language running a slow algorithm will always be beaten by a slow language running a fast algorithm 100% of the time
Nikolaj, the language-based execution time thing was clearly tangential to asdf's point about the time complexity of the algorithm itself. Parallelism is also a polynomial speed-up, so by analogy to your own argument it won't make the algorithm fundamentally faster.
If I may ask, what software did you use to make the lovely diagrams and tables?
Mostly Adobe Illustrator + In Design.
It'll take about an hour.
The function of possible rules given our X pixel rule.
F(x)=2^x (81 choose x) that is our problem.
This is a very exponential function.
So, How do we fix this?
Great Question!
Hi,
Could you share the ipython notebooks/code?
Coming super soon!
That's why ML people so often have incredibly sophisticated coffee habits. They have all the time they need for that while their code is running.
PS this also predicts only the moderately bad programmers should have these
The suspense is killing me... WILL IT GENERALIZE?
This is fucking awesome.
That should be 2^4*(81Choose2) rules, isn't it?
3:40 people picking out stereotypes
Are you that precise with every cup of coffee?
Hahaha nope. Just trying to look like I knew what i was doing.
please do a calculus series too! :D
I think 3blue1brown is about to come out with a really good calculus series!
Welch Labs i also subscribe to him! Thanks!! keep up the really nice animations too (i watched your complex numbers series about 3 to 4 times :D)
if i may ask, what is the name of the ending song ?
why you not pouring water at once..?instead of pouting small small quantity..
I T W I L L G E N E R A L I Z E !
3:42
Although you are bitching about your not-succinct series, I think it's quite succinct since the episodes are, and you can pause and ponder in between the episodes (and during the episodes, I hit pause during an episode a thousand times, since it's so short and succinct, maybe also because I'm tired..) :)
what's that song at the end?
Outro song?
Please please please next: Speech Recognition
Sounds fun! But gosh, how many freaking episodes would that take me..."Last time, in part 234..."
Welch Labs
Maybe combine related episodes next to each other into one.
You used python?
You make videos on the Internet?
Yeah! Internet videos!