Learning To See [Part 13: Heuristics]
Вставка
- Опубліковано 17 бер 2017
- 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/stephencwelch/Lear...
welchlabs.com
@welchlabs - Наука та технологія
How to watch Welch Labs:
-See that pop-up every now and then
-Get Popcorn
-Go 1080p
-Go Fullscreen
Because you deserve that! :)
EDIT: Be sad because it's already over :(
I am an ecologist I go 480, full sound, and a cup of water
Watch the last one first, then the new one!
No, the way to watch is to wait three months, then binge watch them all.
:)
8:55 NO, DON'T YOU DARE TO END THE EPISODE HERE!
9:10 FUCK!
these series of videos are the best I´ve seen, I love them, the only problem is wait for the next one!
Welch Labs is amazing! I love♥️ his videos. Really entertaining, calming, motivative and more importantly: informative!
Keep doing what your doing man!
5 stars for you ⭐️⭐️⭐️⭐️⭐️
Making shit up -> Heuristics
Man, you cracked me up so much. There's literally no other way I could define heuristics ahahah
My friend, you won all my attention and finish the vídeo with "...but we will learn it in the next vídeo!" :D
I don't resist to wait tomorrow although now is late!
Please! I beg you do a video on trigonometry!
The videos you made are awesome!
Thanks for the video! Really enjoying this series.
great video! can't wait for the next one!
Thank you I loved this video so much I pressed like and turned my monitor upside down so I could press it again
This series is gonna make me cry because it's never simple and it's never over but that's computer science I guess
Great series, can't wait for the next video.
I want to start by saying these videos have been awesome, I will be sure to watch the full series. One thing I noticed, the equations seem to be flipped at 8:21 in for the top row.
Guys if you haven't noticed we already saw what curve he is going to use in the first episode, it's just a semi-circle.
always waiting for the next one...
I love these videos, I just wish they din't boost my confidence of knowledge so much unlike my knowledge.
This was awesome stuff, thanks !
My guess for the impurity function is a non linear one : a one that will make your average value for the whole node actually change by a lot when a choice has high impurities, penalizing a lot the node that are even a bit impure (I believe you're going to use a gaussian curve more than a parabola, but not sure)
Question: would quantum computers be able to permit the 'long' method of machine learning (the one with exponential growth)?
I can't handle the suspense!
Random forests do shockingly well for how easy they are to train. At 1 / 1000th the train time of a neural network you get 90 - 95% of the accuracy on some tasks, and superior accuracy on some non-sensory tasks.
Is the answer the normal distribution? If my Probability and Statistics course taught me anything, the answer's always the normal distribution.
Immediately lose faith in statistics.
RandomPanda0 Its not you can see which curve he is usingin episode 1
omgggggggggg i want next part so badly
Why take a weighed average? Why not just use the minimum of both impurities? Since we're classifying all data, an increase in impurity on one side will always lead to a decrease on the other. Finding the node that makes the line clearer by having impurities that are extreme i.e. having the largest difference in impurity seems like a smarter idea.
The minimum won't work because once you have a completely pure node it will say you have perfect purity, even if the rest of the tree is all mixed up.
The sum doesn't work because as you add more nodes the sum will increase and you want it to be a normalized value. So you divide the sum by the number of nodes, which is just the average and has the problem he mentioned in the video where it doesn't work right if nodes have different numbers of entries. For example, if you had a node that separated out a single event on one side and left the rest all mixed up, it would get a high score from a simple sum or average because half of the nodes are classified correctly even though most of the data isn't.
Damn cliffhanger again! Is it about splitting in such a way as to maximise the normalised information gain? (Or entropy?)
Great video again!
I don't wanna wait until next time
I absolutely love these videos, but I wonder. Why are the videos so short and infrequent? These are absolutely amazing, and I would love it so much more if we could get more.
Well, you see, videos take time and effort to make. If the video seems effortless to make, then the video creator is doing a good job, but honestly, making videos like this is not a quick and simple thing to do.
Good quality videos take a lot of time.
All the details and perfection in these videos require a lot of time!
Although I would love to see them come up more often, I wouldn't like them being rushed and unfinished.
He's doing an excellent job, I love it! :D
Infrequent? They have been coming out every two weeks! So glad!
The weighted average impurity for every split is 2/5 because 2/5 of our training examples belong to the minority class (yellow).
I love this videos!
Fastantuc Channel I am a new subscriber from Colombiano, I love the Ciencia, Physics, Mathematics and Química xD
My heuristic answer for the problem: get rid of linearity!
Good call!
I'm in so much suspense
The hype!!!
yaay more videos
You should put "Machine Learning", gets a lot more views. Great videos!
Time for some educated guesses!
Part 13, another cliffhanger ;_;
Great video, as always!
What IDE do you use for your python code?
Jupyter notebook! Thanks!
I guess instead of taking the total impurity of the whole classification, but rather the impurity in each labeled group and choose the classification that contains the least impure group within.
Yep!
how long do these videos take to make?
and is it realistic for you to upload more often?
+Alex Gregory
He gotta do research. Make a script. Polish the script. Record the audio. Make time lapses for visualizing. Make computer graphics for visualizing. Edit the video and upload it.
That's a lot of work.
I can tell that these videos are a lot of work, I was just wondering how long it took to make one
and I also know there are reasons to upload less often than you can (namely consistency)
I wonder if he works full time on the side. He should set up a Patreon account.
(Next Time) , Come-On Seriously
mistake at 8:20, some of the equations dont match up the IR and IL.
No part 14 ? Damn it, I want my drug!
@1.51 since the bracnhes are classified both as -(the classifcation erros are a summation of both).. correct?
I see comments complaining about there being less info in videos and more examples/things to experiment with . But in my opinion, this is exactly the approach which is required in order to build a scientific understanding and intuition for something. It is these examples that made me feel like i was the one discovering all of this myself step by step and it kind of simulates what actually goes inside the minds of great scientists and researchers. I always thought this is the right way to learn something or get someone interested/intrigued in something. So well done Welch Labs and keep making videos like you do! They are amazing! :D
Thank you!
I think you switched the I_Totals for x_1 and x_2 at 8:22. Maybe put an annotation because this confused me a bit. Great work though :)
Sorry about that - can't believe I missed it!
damn.. that cliffhanger tho
Really enjoy these videos, but the music during the talk makes it very hard to follow
at 8:36 you put the caculations for I[total of 1] under I[2] and vise versa for I[1].
Noooooooooo! That would drive me insane! I think that I[3] could be off as well. That is so something I would do.
Damn! I can't believe I missed this! I really need to do some type of peer review on the next series.
Where is your Patreon ?
Coming later this year!
Any other way to support you in the meantime?
I do have a book for sale: www.welchlabs.com/resources Thanks for watching!
Square mean error?
Ochi chernye
Entropy!
:)
I came as fast as I heard
you sound like ian stewart
WATCH THIS ONE SIMPLE CRAZY TRICK THAT WILL MAKE AI SMART
8:35 the equations are in the wrong places.
MLE?
but I wanna know nowwwwwwwwwwwwwwwwwwwww
My intuition: you use a normal distribution instead of a triangular one.
As an approximation to the binomial distribution?!? I could totally see that!
Iustinian Constantinescu Nope he uses a semi-circle as can be seen in episode one
The impurity heuristic function _I(p)_ is not a probability distribution!
While we could plug in parts of the formula that describes the shape of the normally distributed probability density function (the bell curvy thing) to kinda make it work (and maybe even get decent learning results by luck), I can't think of any theoretical justification for why that would make any sense. So that should make you feel suspicious about that intuition.
(You're right about the rough shape of it though. We're looking for a concave function, and the middle part of the normal pdf is concave. ;) )
nibblrrr Thank you!!
*is this bold*
*woah this is pretty cool*
_amazing_
noooo not again
After each of these videos I wonder what have I learned? -That the next episode will be really important
Try to put more info in one episode please. Like all you said in this one is all examples seem equally bad.
I feel bad I understand nothing.
In the previous video, you've said "Let's test our new strategy on real data. But before we do, let's consider how our new strategy might perform." But you didn't test that strategy in the previous video. You didn't test the strategy in this video. I doubt you will test that strategy in the next video, because in this one you've already changed the strategy. So why say "let's test our new strategy" in the first place?
The music is distracting. It might appeal to those who are just pretending to be smart, but not to those who are really trying to follow.
Otherwise I'm super impressed with the production.
Are we ever gonna get to the point? I really like the videos, but every single time it ends, I'm like "wtf? That was just an introduction".
If it were the complex number series, it would have ended now, maybe the heuristic that guides its path to an end is not admissible? PS: Already unsubscribed.