I like your presentation, wit and pacing. But black backgrounds and white text is just killing my eyes. I do realise that the majority of artists like black environment. Just wanted to point out that a minority exists as well. It would be great if all the software would be switchable to a light mode for accessability reasons.
His style reminds me of codebullet, I actually was a little confused because I clicked on this video subconsciously and I thought it was him for about the first quarter of the video.
The production quality on this is insane. When i clicked i thought it was 596 k subscribers not 596. Then i saw that when you made this u had 38. THATS CRAZYYYYYY. Subscribed and im probably gonna bingewatch your videos now. Your style reminds me of code bullets.
Honestly I think you can go above an beyond with res nets similar to resnet50 or alpha zero's net arch and then add a multilayered perception, it helps a lot with when compared to a fully connected layer(Alexnet). Truthfully speaking, just throwing more layers and neurons increases the accuracy, but also increases the chances of overfitting.
Awesome video! I did a 1st year engineering project that looked at different solutions to MNIST, our most accurate model was very similar to Alexnet but with some parameter / layer changes that let it get up to 99.7%. However, there are much less complex models that can reach the 99% mark.
Wow! This is incredible! I assumed because of your thumbnail and while I was watching the video that I was watching someone with 100k+ subscribers yet I've come to find you only have 385... I'll add another to start the way! I'm not much of a programmer, but this video and its style is incredible! I'm waiting for the video on Sunday ;)
Two linear layers without any non-linearity as activation function in between is exactly the same as one linear layer. Think of a linear layer as (matrix) multiplication: out=weight1×(weight2×input) is the same as out=(weight1×weight2)×input . So the first two models are almost the same, just written a bit differently.. Thid explains why they perform the same..
I'm 120% sure is a 7, no one can think is another thing if you don't put a 7 at side for reference. Humans have the same problem trying to discern between a 0 and a O if you don't have a context.
Hey, first i wanna say I really like the style of the vid and its cool to be here when the channel is still small. Can I ask how did u learn to code, and how long have you been at it for?
Thank you so much! :D I started learning to code I think maybe 8 years ago. I mostly just watched tutorial videos and I built little projects. I'm still not great at it lol, but I have been slowly improving.
This is the first video of yours that was recommended to me on TH-cam, and it was the best thing that happened to me today. Keep it up, bro. I enjoy your videos, but don't forget when you have more than 1 million subscribers 👻
Yo boi, dont you stop uploading these type of videos, you will one day reach a high feat, and 10k by the end of 2023 i wish i could donate but i dont have anything one me rn, keep going g
Wow this is amazing, may not seem like much now but this is top notice entertainment give it a few years and I have no doubt you’ll have quite a bit of subs (50k-100k) then it’s just growing after that
The quality of this video is excellent. As an ML engineer, I suggest focusing on the F1 score on a previously separated testing dataset (There will be 3 divisions - training, testing, and validation). It would be best if you also created a graph from the history of training the model including the validation dataset. You see accuracy is not the one that we seek, we see how much pure the training has become. Also, it is essential to explain why you are applying a model, though most of the time, it's difficult to understand.
The problem with machine learning for symbol recognition is that we do not only recognize symbols from their shape, but also through the context. For example, that half-drawn zero, we would not be able to know that it was a zero unless we knew that it was supposed to be a number. And I would also argue that if you asked humans about that 1 you drew, you would likely have a large percentage guessing a 7. Personally I think drawing ones with that hook at the top should be frowned upon in general. It should just be a single line, like l or I.
What is backpropagation learning algorithm? Backpropagation, or backward propagation of errors, is an algorithm that is designed to test for errors working back from output nodes to input nodes. It is an important mathematical tool for improving the accuracy of predictions in data mining and machine learning.
I wrote my own neural net in C#. With 1 hidden layer of 255 nodes it got 97.6% accuracy. With 2 hidden layers of 255 nodes each it also dropped to 92%. No idea why.
In the beginning you implied that the error was based on the distance between what it thought the number was vs what it actually was Wouldn't a better way to calculate the error to base it on shape? a 3 doesn't look like a 1 but the error (when based on values) is only 2
Great, thanks for the video. I personally like the details of implementation. If it's possible to create a video that would serve as a walkthrough of practice basics, that would be great learning material.
I guess you have made visualization of dummy model a little bit wrong. There is no single hidden neuron. It has just flattened layer of input and dense connection with output layer, no hidden. also single neuron doesn’t make sense and can not reach 90%+ 😅. But anyway good vid
Amazing video! Will you make tutorials on how to do something like this? It would probably be the most informative tutorial on Machine Learning I've ever seen
In the second network I think the accuracy is low because the module is OVER FITTING I think if you decrease the number of nodes and find a normal threshold number the accuracy increases
Agreed. It's also important to note the database includes digit "1" in the form it is used in English speaking countries e.g. the US - a single straight line. Testing ist with a European style "1" (2 lines) will result in being not recognised as a one. Because the tester writes the digit 1 the way the system is trained with as a 7. IMHO not the AI but rather the tester is the issue here :-)
I'm studying to become a software architect so I don't understand anything about AI. Which linguage are you using? Are your program being executed on che GPU?
the database includes digit "1" in the form it is used in English speaking countries e.g. the US - a single straight line. Testing ist with a European style "1" (2 lines) will result in being not recognised as a one. Because the tester writes the digit 1 the way the system is trained with as a 7. IMHO not the AI but rather the tester is the issue here :-)
I was wondering if there are people who actually watch these videos, these types of videos are trending in the most watched ones Nobody can interfere in anyone's life!definitely, probably, I suppose, I guess there's a chance I'll be a trend someday?
That's the problem, you can't really analyze the created net and figure out why it recognized a very well drawn 9 as a 3. The problem is even more of a bigger deal, because even pixels that have nothing to do with the number can screw up the net. Like add one white pixel on the edge and watch your net failing. Called Pixel Attack.
But actually can someone tell why a model would predict a 9 to be a 3, even though it accurately predicted a 3 right initially in the same array of numbers. Is it memory issue? Not an expert, but this triggers interest. Great content.
Okay, just check. I though it I set all the learning rates to be equal, but turns out that no. For the first and last model, the learning rate was 0.1. And for the other models it was 0.05. Thank you for pointing that out :)
Agreed. It's also important to note the database includes digit "1" in the form it is used in English speaking countries e.g. the US - a single straight line. Testing ist with a European style "1" (2 lines) will result in being not recognised as a one. Because the tester writes the digit 1 the way the system is trained with as a 7. IMHO not the AI but rather the tester is the issue here :-)
![เขามัทรี - เอ็กซ์ ศุภกฤต - จอนนี่มิวสิค [ Official MV ]](http://i.ytimg.com/vi/AS9ufKuQIVs/mqdefault.jpg)
Thank you so much for watching☺! Make sure to check out the fundraiser💚: gofundme.com/raising-10000-for-diabetes-by-the-end-of-2022
2022 already ended
@@millionare5446 NAAAAAH REALLY
Diabete
I like your presentation, wit and pacing. But black backgrounds and white text is just killing my eyes. I do realise that the majority of artists like black environment. Just wanted to point out that a minority exists as well. It would be great if all the software would be switchable to a light mode for accessability reasons.
Keep it going. Your production quality and also kind of self ironic style is fantastic.
Thank you so much! :)
His style reminds me of codebullet, I actually was a little confused because I clicked on this video subconsciously and I thought it was him for about the first quarter of the video.
@@DanksterPaws def codebullet inspirrd
Yeh feels like style has been stolen from code bullet. Still huge effort making this kind of videos
This deserves waaaaay more likes and views
EDIT: wow already 256k views and 7k likes did not expect that in such a short amount of time
:)
True
DEFINITELY
The production quality on this is insane. When i clicked i thought it was 596 k subscribers not 596. Then i saw that when you made this u had 38. THATS CRAZYYYYYY. Subscribed and im probably gonna bingewatch your videos now. Your style reminds me of code bullets.
Thank you 💚
Honestly I think you can go above an beyond with res nets similar to resnet50 or alpha zero's net arch and then add a multilayered perception, it helps a lot with when compared to a fully connected layer(Alexnet). Truthfully speaking, just throwing more layers and neurons increases the accuracy, but also increases the chances of overfitting.
Also thank you so much for linking the code that’s awesome!!! And very very appreciated!
Of course!
Awesome video! I did a 1st year engineering project that looked at different solutions to MNIST, our most accurate model was very similar to Alexnet but with some parameter / layer changes that let it get up to 99.7%. However, there are much less complex models that can reach the 99% mark.
Oh nice! 99.7% that's high!
Wow! This is incredible! I assumed because of your thumbnail and while I was watching the video that I was watching someone with 100k+ subscribers yet I've come to find you only have 385... I'll add another to start the way! I'm not much of a programmer, but this video and its style is incredible! I'm waiting for the video on Sunday ;)
Thank you so much man! You made my day :). I'm still editing today's video but it should be uploaded later today. Let me know if you like it 💚
@@Green-Code Man u growin fast, already 1k subs
6:53 wow, it’s interesting way to show accuracy, never seen it before
Absolutely amazing video. Great presentation style. Quality is above 99% 😊
Two linear layers without any non-linearity as activation function in between is exactly the same as one linear layer.
Think of a linear layer as (matrix) multiplication: out=weight1×(weight2×input) is the same as out=(weight1×weight2)×input .
So the first two models are almost the same, just written a bit differently..
Thid explains why they perform the same..
Oh shit thank you for that
5:56 honestly that does look way more like a 7 than a 1 😂
Yep, my drawing skills are not the best 😂
@@Green-Code haha you did great! keep up the AI videos btw they have so much potential
I'm 120% sure is a 7, no one can think is another thing if you don't put a 7 at side for reference. Humans have the same problem trying to discern between a 0 and a O if you don't have a context.
WTF! I just saw you only have 100 subs?! With this quality you deserve more. You have my sub.
Hey, first i wanna say I really like the style of the vid and its cool to be here when the channel is still small. Can I ask how did u learn to code, and how long have you been at it for?
Thank you so much! :D
I started learning to code I think maybe 8 years ago. I mostly just watched tutorial videos and I built little projects.
I'm still not great at it lol, but I have been slowly improving.
@@Green-Code can you give some ideas of what to do in python? I wanna learn to use it but iam not sure what to do with it
*It is a crime that you have 4000 followers.* You deserve 4 million 🤯
The fact that I studied AI at university and this video was more understandable than my professor is crazy
Im in High School, and I have learnt more from my classmate than from my Physics teacher
This is the first video of yours that was recommended to me on TH-cam, and it was the best thing that happened to me today. Keep it up, bro. I enjoy your videos, but don't forget when you have more than 1 million subscribers 👻
Wow thank you dude. I won't forget :)
I had to rub my eyes to make sure your sub count really said 386 subs and not 386k subs
wildly underrated, excited to see more content
Thank you :D
Been playing with mnist lately too, cool to see your results. Looking forward to more vids :)
Oh, that's so cool. Pytorch or Tensorflow?
Yo boi, dont you stop uploading these type of videos, you will one day reach a high feat, and 10k by the end of 2023 i wish i could donate but i dont have anything one me rn, keep going g
Thank you for the kind words :)
I have also had the pleasure of completing the CNN course on Kaggle,
This video came at a great time, also best of luck with your channel!
Thank you so much! I don't really know what course that is, but I hope it was interesting :)
Wow this is amazing, may not seem like much now but this is top notice entertainment give it a few years and I have no doubt you’ll have quite a bit of subs (50k-100k) then it’s just growing after that
Thank you for the kind words :)
Great stuff! Looking forward to seeing the rest of your work
Nice video format. Keep it up!
The model sees numbers I can't see XD
You and me both XD
The quality of this video is excellent. As an ML engineer, I suggest focusing on the F1 score on a previously separated testing dataset (There will be 3 divisions - training, testing, and validation). It would be best if you also created a graph from the history of training the model including the validation dataset. You see accuracy is not the one that we seek, we see how much pure the training has become. Also, it is essential to explain why you are applying a model, though most of the time, it's difficult to understand.
Thank you for the tip :)
i just want to have a comment on your channel from the start of your journey to 100k , good luck you got it with an awesome content
Thank you so much :)
Now I know why my drawn numbers don't get recognized in games
It's unbelievable how with your video I learnt a lot about AI. Thank u so much!!!
Happy to hear that!
1:20 where did you get the model of my brain during exams from
The problem with machine learning for symbol recognition is that we do not only recognize symbols from their shape, but also through the context. For example, that half-drawn zero, we would not be able to know that it was a zero unless we knew that it was supposed to be a number. And I would also argue that if you asked humans about that 1 you drew, you would likely have a large percentage guessing a 7. Personally I think drawing ones with that hook at the top should be frowned upon in general. It should just be a single line, like l or I.
bro imagine the confusion with Il| and the 1 symbol
also unfortunately that won't happen (look at this symbol here --> 1
but yeah everything else there is right
Your really fun to watch, I’m to dumb to understand anything ur saying but it was a fun watch regardless
What is backpropagation learning algorithm?
Backpropagation, or backward propagation of errors, is an algorithm that is designed to test for errors working back from output nodes to input nodes. It is an important mathematical tool for improving the accuracy of predictions in data mining and machine learning.
... 3 - 6 - 9 - Tesla Numbers its realy enough to KNOW.
5:57 I honestly thought that was a seven too
This is such a well made video! You deserve way more subscribers. Keep it up!❤🎉😊
Thank you for the kind words! New video today :)
I wrote my own neural net in C#. With 1 hidden layer of 255 nodes it got 97.6% accuracy. With 2 hidden layers of 255 nodes each it also dropped to 92%. No idea why.
Wow, that's so interesting. Maybe it's overfitting?
In the beginning you implied that the error was based on the distance between what it thought the number was vs what it actually was
Wouldn't a better way to calculate the error to base it on shape?
a 3 doesn't look like a 1 but the error (when based on values) is only 2
Yeah, you're right it calculates the error based on the shape :)
It was just an easy explain what I was doing
At the thumbnail, you should swipe the screen back and forth and you would probably see an illusion of that curve
Model 2 doesn’t improve from model 1 because you are only using liner layers, try changing the activation function to relu
Great video! Learned something new and would love to learn more!! Thank you!
Hello, why checking for accuracy and not f1-score? Thanks.
I just thought it would be easier :)
Great, thanks for the video. I personally like the details of implementation. If it's possible to create a video that would serve as a walkthrough of practice basics, that would be great learning material.
Thank you for the suggestion! I might do that in the future
Great video man, you make it interesting all the way until the end! I also checked your blog, nice article about science and code! +1 sub and share
Thank you bro, I appreciate that! I haven't touched the blog in a long time, but I'm very glad you enjoy it :)
thank you very much for giving the code out to everyone
Woow! Bro you explained very well! Amazing job 👍
my homework on special mathematics subject in computer engeneering was literaly this :)
really good content oh and IM UR 100TH SUB LESGO
Let's go!!!
Thank you so much, now I will build something so smart that politicians can be replaced
Okay, let me know how it goes
Great video keep it up!
I guess you have made visualization of dummy model a little bit wrong. There is no single hidden neuron. It has just flattened layer of input and dense connection with output layer, no hidden. also single neuron doesn’t make sense and can not reach 90%+ 😅. But anyway good vid
Yep 😅! You're absolutely right :)
Amazing video! Will you make tutorials on how to do something like this? It would probably be the most informative tutorial on Machine Learning I've ever seen
Possibly :) (but not at the moment)
good job dude, you are keeping it straight forward
now do the alphabet
Got lost in thought at "neuron." I don't have many of those left.
what tool did you use to visualize the convolutions?
In the second network I think the accuracy is low because the module is OVER FITTING
I think if you decrease the number of nodes and find a normal threshold number the accuracy increases
Fair enough! Thanks for the suggestion :)
love the video man. Looking forward for future uploads
Thank you! More to come!
Looks like you gained like 200 subs in 10 days. I guess you'll gain even more if you keep this up.
I'll guess we'll see :) New video this Sunday!
ok, but what can neural networks be used for besides recognizing images?
Watch the new video I'm going to post on Sunday and you'll see :)
Thanks for the video, this was really interesting!
Very cool , i need to learn Machine Learning , Neural Networks too :)
Before watching this I thought he was going to train the models with the numbers from a bunch of different fonts...
This is awesome man ... 🤙🤙🤙Keep rocking 🤙🤙🤙
Again amazing video! You should make more videos!!!
At 5:06, you CANNOT tell me the bottom left one is a f***ing 8
Cool Video :) reminds me a little bit of Code Bullet
Could you please make video about ML Alexnet how to code etc?
Where did you learn it from?
There's some great tutorials out there on Pytorch and Tensorflow that are completely free. Check Daniel Bourke out!
this video idea is not original and i saw it before, but still amazing realization!
6:00 dude that is literally just a 7
My drawing skills are not the best
Agreed. It's also important to note the database includes digit "1" in the form it is used in English speaking countries e.g. the US - a single straight line.
Testing ist with a European style "1" (2 lines) will result in being not recognised as a one. Because the tester writes the digit 1 the way the system is trained with as a 7.
IMHO not the AI but rather the tester is the issue here :-)
Mission impossible: Teaching AI to read doctor's handwriting
😂😂😂
I guess the second model was progressed wrong. I get 97% accuracy on the test set with this architecture.
Yeah, could be maybe there's some sort of bug on my code
Nice viedo, you got me interested!
I'm studying to become a software architect so I don't understand anything about AI. Which linguage are you using? Are your program being executed on che GPU?
I'm using PyTorch (python) and yes my program is being executed on the GPU :)
6:07 for the hard 1 I would call that a 7
Yeah, fair enough my drawing skills are not the best
@@Green-Code it’s ok sometimes my 7s look like that
the database includes digit "1" in the form it is used in English speaking countries e.g. the US - a single straight line.
Testing ist with a European style "1" (2 lines) will result in being not recognised as a one. Because the tester writes the digit 1 the way the system is trained with as a 7.
IMHO not the AI but rather the tester is the issue here :-)
I was wondering if there are people who actually watch these videos, these types of videos are trending in the most watched ones Nobody can interfere in anyone's life!definitely, probably, I suppose, I guess there's a chance I'll be a trend someday?
5:57 ok even I thought that was a 7
That's the problem, you can't really analyze the created net and figure out why it recognized a very well drawn 9 as a 3. The problem is even more of a bigger deal, because even pixels that have nothing to do with the number can screw up the net. Like add one white pixel on the edge and watch your net failing. Called Pixel Attack.
But actually can someone tell why a model would predict a 9 to be a 3, even though it accurately predicted a 3 right initially in the same array of numbers. Is it memory issue? Not an expert, but this triggers interest.
Great content.
Neural nets are kind of a black box, so I don't really know :)
Like optical character recognition technology
Same here
Now we can tell Mason what the numbers mean.
better to split to 70% training and 30% testing, just my opinion
Why don't they put the bar on the seven ? All the seven look like ones
You fail to mention that your dog is named "diabetes"..... :>
6:38 how is this at all exactly 10 minutes and 10 seconds? It's not even exactly 10 minutes and 30 seconds
Yep
For AI is better to use TPU (TensorFlow Process Unit) if you use TensorFlow in your projects
Didn't know that, thank you😊
@@Green-Code No problem mate 👍💪
The thing is though that some of those numbers would fool a human such as the 1 that looks like a 7.
True
Amazing video! Did you use the same learning rates for all the models?
Thank you! Let me check :)
Okay, just check. I though it I set all the learning rates to be equal, but turns out that no. For the first and last model, the learning rate was 0.1. And for the other models it was 0.05. Thank you for pointing that out :)
great intro to neural nets :D
How or from where can i learn this
LOL, i did exactly the same thing when i was doing Computer Engineer!
6:10 Even I thought that was a 7
Agreed. It's also important to note the database includes digit "1" in the form it is used in English speaking countries e.g. the US - a single straight line.
Testing ist with a European style "1" (2 lines) will result in being not recognised as a one. Because the tester writes the digit 1 the way the system is trained with as a 7.
IMHO not the AI but rather the tester is the issue here :-)
Getting Code Bullet vibes here
I subscribed to you because you gave code
U have
Maybe one day lol
now train a Vision Transformer
Is it possible to get 99.99%
I guess it isn't. It probably depends on how good your model is and the quality of your testing dataset :)
CAPTCHA isn't safe anymore-
Great video ma man ,, i learned a lot
50% of the c0mments: "kEeP iT uP!!!!1!!eleven!"
The other 50% of the comments: hoio shingienyio hoio
Nothing says "AI" more than jQuery of a web app in splashart 0:05
Yes