Convolutional Neural Networks from Scratch | In Depth

[@jackfarah7494]

I have been researching CNNs for about a month now. Every video i watch, i end up more confused and no answers. I cant express how grateful i am of this video. Thank you so much for this great content and educational information. Keep it up man!

[@far1din]

Thank you my friend. Glad you got some value out of the video! 💯

[@maxave7448]

​@@far1dini absolitely love how this looks like a 3Blue1Brown video but doesn't throw a bunch of numbers and terms at the viewer nonstop. This is great for beginners!

[@notsoclearsky]

Bruh I can't thank you enough, this is some gold tier education literally. Keep up the good work

[@mendezzzzz123]

This is amazing, thanks, nothing better to understand this abstract concept, just visualizing it

[@HoshRampageZA]

Wow! I finally feel like I understand neural networks including the math at every stage. I have never seen a complicated math formula broken down so simply and elegantly. You are an excellent teacher. Thank you for this video. Subscribed.

[@chinmaythummalapalli8655]

I racked my brain for hours and couldn't figure out why the features' maps aren't multiplying after each layer and this video just helped me realize they become channels of images , it helped me relax and I think I can go downstairs for dinner now.

[@far1din]

Glad it helped! 😄

[@sachink9102]

WooooooW ! I am speechless man ! you are THE genius

[@nettwrrk]

Can't express my gratitude, albeit here I am. Everything is shown very detailed, explained accurately and understandably. Keep up the good work.

[@naveens482]

I have learnt CNN on many platforms but this video is the one that thought me exactly what i need .Need more videos like this

[@paedrufernando2351]

you helped clear the finishing clincher for me in the worlkd of AI..cant thank you enough

[@tebs1989]

This is the most clear and exceptional video explanation about CNNs that I ever seen so far. Thank you so much!

[@andyh3970]

The single best explanation I have seen 15/10

[@prashantkesharwani9205]

Thank you so much for creating such an informative content. Keep it up, your channel is so much underrated!

[@卞正-s5y]

I have learned a lot from this video. It is beneficial for people like me who don't have studied CNN at all and want to learn something.

[@im-Anarchy]

Sankyo zo munch vor this veautifull, vuonderfull ,amazing video. Arigatoya!

[@far1din]

Thank you bradder :D

[@im-Anarchy]

@@far1din Bradder??? do you mean: A word for a particularly unattractive female, usually with reference to a slightly deformed smile.

[@far1din]

😂😂 no my friend. Just saw this definition on urban dictionary. it’s a typo for «brother» 🥇

[@ViralKiller]

Amazing, didn't understand crap until you explained it with images...please make more

[@imadboukhari8033]

Great video man better than anyone out there thank you

[@SelfBuiltWealth]

this is a very unique and underrated explanation!beautiful work thank you so much❤

[@JamieTx23]

Excellent video! Thanks for taking the time and breaking it down so clearly.

[@far1din]

Very welcome!

[@jayeshkurdekar126]

You are a gold 🥇 professor..wish I was a billionaire..would have gifted to you for you sheer clarity

[@far1din]

Haha, I hope you become a billionaire one day 💯

[@Blackprogger]

Thank you for this great explanation! It couldn't be explained any better! Very nicely visualized and explained step by step! The best explanation of CNNs I've seen so far! Thanks!

[@manishsoni8806]

Awesome Explanation 😍😍

[@hewramanwaran6444]

Great Explanation. Thank you very much.

[@let1742]

thank you! this is the clearest explanation i've seen, i hope you will continue to produce videos of this kind!

[@pappo-nc5yh]

Great video and very clear explanation, thanks!

[@samuelbrouwer430]

this could not be more clear thank you

[@mateokladaric]

Finally someone who doesn't just say "it convoluted the image and poof one magic later it works"

[@lushbeard]

This is fantastic level of explanation

[@SamuelMoyoNdovie]

What an explanation man 🫡

[@Ivan-fz3ou]

Awesome work! This gave me a new insight and understanding of CNNs; the intricacies and math of how it works.

[@birajkumarkaranjit7259]

very well explained

[@immohobot9288]

Nice explanation. It was really helpful. Thanks.

[@tamurhaq]

Excellent content. You've made this keeping in mind the viewer's intuition. Keep making more just like this. ❤

[@rangilanaoermajhi1820]

Brilliant! Looking for more visual representations! In the final layer of softmax, if you can also explain how its going to 7 using the learnt parameters. Big thanks 👍

[@franciscobrizuela766]

Thank you! Now I'm one step closer to finishing a model for hw :)

[@far1din]

You can do it!

[@ratfuk9340]

Awesome, this cleared things up for me. Thanks!

[@satellitesabunim]

Excellent video.

[@joseluisdiaz233]

An amazing job, thank you for your time and for sharing

[@danny2704]

Pretty like this visualization !!

[@SelfBuiltWealth]

beautiful explanation❤

[@RAHUL1181995]

This was really helpful....Thank you so much for the vizualization...Keep up the good work...Looking forward to your future uploads.

[@samruddhisaoji7195]

Thank you! your explanation and animations were very helpful!

[@jaybhatt6775]

wow.amazing illustrations!

[@OnlineClasses-rs5yf]

Great work... Much better than college professors

[@AjinFrankJ]

wow this is a gem

[@Temuei]

Thanks, this video very easy to understanding for me

[@Tezla0]

Really good content. You deserve more subscribers and views.

[@kemaldursun8192]

thank u man it's great content and helped me so much

[@HiepNguyen-bw6dj]

thank you so much! good explanation!

[@bdeceulaer]

Brillant visualisation and explanation! Your video clarified to me in minutes the difference between a convolutional layer and a fully connected one, the meaning of stride size, max pooling and activation function. What is the impact of different activation functions? I assume weights, biases and filter values are determined iteratively during training. Would be great to have a visualisation video of that training phase for this same image recognition example.

[@far1din]

Hi Bart, and thank you! 1. The activation function. The activation function induces non linearity to network. I highly suggest you watch Andrew NG's video on this as he explains the mathematics behind. I have referenced it below for you! :) An intuition I once heard that stuck with me is that you want these neurons in the hidden layers to fire. When using activation functions such as ReLU, this is exactly what happens. If the calculated value goes below zero, the "neuron" in the next layer is set to zero. 2. Weights, biases and filters are set iteratively? You are correct that the weights and biases are set iteratively during the backpropagation, but the filters are fixed/predefined. I will try to make a video on the complete training process for the next video! :) ref: kzbin.info/www/bejne/hJyyp5KhbNdppNE

[@far1din]

kzbin.info/www/bejne/aJ_Vo61_rcScask

[@khayyamnaeem5601]

Amazing content!

[@sajanphilip8221]

Best Explanation ever

[@boklausen9583]

Brilliant explanation!! - thank you so much for sharing! Now, what is the magic (or heuristics) behind defining the various kernels and pools (sizes, strides and contents)?

[@far1din]

It’s an iterative process. Trial and error until you get the best results. Andrew Ng actually talks about choosing hyperparameters in this video: kzbin.info/www/bejne/Z6jEeZ-mgM6Br7ssi=saG0hYPuKg5yHiji

[@boramin3077]

Great explanation!

[@XenoZeduX]

Great

[@debraheric2308]

Wow such great content. Subscribed!

[@peterpan0201]

This is actually very good!

[@Number_Cruncher]

You nicely explained the action of each layer. I wonder if there is an interpretation of the visuals that where seen in the intermediate steps. Or, it would also be nice to see, how the filters evolve from random to there trained configuration. Can the values of the filters be interpreted somehow? I think of edge detection, gradients or something similar.

[@far1din]

Thank you! There are no exact interpretation of the values within the filter that I am aware of. Please comment below if there is. What can be seen, from this video and others is that the first layers will detect the shapes, and the deeper layers will interpret more complex features. Although this is not a mathematical proof, you can see this effect by for example creating three sets of the same model, initializing different weights for each of the three models and training the model on the same set of data. You will notice that the filters converge to different values, but the outputs for each layers are somewhat the same. You start with some form of edge detection and move on to more complex features that in my opinion are hard to identify, at least for handwritten digits. I will try to make a video visualizing the training process so that this effect can be seen! ☺ Here is a video by Andrew NG explaining what deeper layers are learning: kzbin.info/www/bejne/eZnSh2iebNmqa6M

[@far1din]

kzbin.info/www/bejne/aJ_Vo61_rcScask

[@louissimon2463]

this is excellent, thank you

[@imotvoksim]

Very thorough and great visualizations!

[@eneadriancatalin]

One mention: 9:14 the sigmoid function is 1/(1+e^(-x)) and your x is already -7.36 so it will be 1/(1+e^7.36), that's almost 0 (0.000485425106)

[@far1din]

Yes my friend. I had to «scale up» the pixels in order for them to be seen.

[@chinedudimonyeka2856]

Thank you

[@overtrust7143]

Awesome

[@nelsonvanduin2583]

Sick!

[@ButcherTTV]

great video!!!

[@drc9313]

Thanks for the excellent explanation. One suggestion: please remove the background music. It is distracting

[@doctorshadow2482]

Thank you for the nice visualization. Two points: 1. You promised in depth explanation, will it follow? In this video you don't explain where from you take these filters/kernels; in depth explanation doesn't assume something "just given", I need to understand where from to get it and how exactly. 2. There are tons of videos on youtube on this topic, it would be nice if you make a difference, explaining, for example, how all this could work with shift/rotation/scale of the image. Nobody covers this.

[@khvnp1l0t]

In the output layer, is it just the highest value after the calculations that makes the prediction? Wonderful video by the way, this has cleared up a lot of questions for me in general about how a CNN works!

[@tobiaspucher9597]

amazing

[@DmitrievAlexander]

First, thank you really very much! Question: is filters are 'generally random' and then 'trained' through feedback, means that we don't really know why the system generally recognise the image? Why it detects this image like '7'? Am I right? (But again, thank you very very much, I'm visual person, and math converted to images explained everything crystall clear!)

[@far1din]

That is correct. This a really really small model, and it has 1138 trainable parameters. Bigger models like resnets have tens of millions of trainable parameters. There is no way, atleast as of today, where a single person or a group of people can pick and choose/guess what number to add to the filter. Although it should be possible :P The probability for this is almost zero. However, we have a feedbackloop where we start of randomly and train the model with backpropagation. I will link some videos below. You can almost think of this as regression. In linear regression, you start with a scatter plot of point, but you don't know what function will give you the best fit line: - A line has the formula ax + b. You do some math, and you can solve for constant a and b. Add more points, and you'll most likely get different a and b values. - For this model, you do backpropagation and solve for 1138 parameters. You change the training images, and you'll most likely get different filter values and biases. I hope this made sense! :) Backpropagation in neural networks by @3blue1brown: kzbin.info/www/bejne/f53KZJp9mtyEa7c Backpropagation in convolutional neural networks by me: kzbin.info/www/bejne/sGrLe62aqq2HpcU

[@adrianhochla3664]

I really like this!

[@hchattaway]

awesome video that really helped understand the inner workings! I've always wondered about each hidden layer... Is it true that for each kernel used there is one of these layers? And when using a tool like PyTorch are there just standard kernels that are used for pulling out features? Is there control over the make up of those kernels? Also, I can imagine, depending on the nature of the images being trained, that custom kernels could be created to best pull out features for a particular data set? Thanks for the awesome work!

[@far1din]

Thank you my friend! 1. I’m not sure what you are referring to, but each convolutional layer will output an activation layer. If this activation layer is «sandwiched» between the output and the input, it’s called a hidden layer. A convolutional layer usually have many filters within. Each filter will return one activation matrix. The activation layer is basically a combined term for all the «activation matrices». Also, some publications referr to the pooling layer as an individual layer while others don’t. In this video the pooling was not considered an individual layer. 2. There are different methods for initializing the kernel/filters, but here we just used the default initialization which is called «glorot uniform». After the training process, you can save the weights and reuse them as you would like. You could also use different initializations or try your own custom ones in pytorch etc. See reference 1. 3. That’s correct. The convolutional neural network shown in this video will be excellent at predicting handwritten digits, but would do poorly in detecting for example handwritten letters. However, the weight’s can be reused in order to train a model which detects handwritten letters. See ref 2 where Andrew NG explains how to implement open source models. Hope this answered your doubts! 🚀 Reference 1: discuss.pytorch.org/t/initialize-weights-of-convolution-layer/52672 Reference 2: kzbin.info/www/bejne/mXepppKVosiif9k

[@FelLoss0]

very well explained!!! new subscriber here :)

[@Deepak-ip1se]

Very nice video

[@manfredbogner9799]

Sehr gut😊😊

[@far1din]

Danke 🥺

[@MalamIbnMalam]

Is there a website where we can solve sample problems pertaining to CNN and RNN?

[@Satrix1689]

Hi Im new to this , may I know how do you get the bias term at the first layer and also the bias term and weight at the output node?

[@far1din]

Hello, the bias start of as 0 and it changes through training (backpropagation). The weights are generally assigned "randomly" and they also change through backpropagation. I will link some videos on backpropagation. I did not include the bias term in my video as I was focusing on the weights, but 3blue1brown did and he has a really good video on this topic. Although he is explaining regular neural networks, this might help you get some clarity as for the bias terms and the final fully connected layer :D Backpropagation in neural networks by @3blue1brown: kzbin.info/www/bejne/f53KZJp9mtyEa7c Backpropagation in convolutional neural networks by me: kzbin.info/www/bejne/sGrLe62aqq2HpcU Training convolutional neural networks by me: kzbin.info/www/bejne/aJ_Vo61_rcScask

[@rotemlv]

I'm curious why you used the sigmoid function in particular - did you get better accuracy using it in this model than with 2 ReLUs?

[@far1din]

Hey Rotem, I only used sigmoid to show that different activation functions could be used. I didn’t want anybody watching to think that ReLU is the only activation function. This video was made for educational purposes, and I didn’t think much of the accuracy as it was above 90% On the sidenote, I was also supposed to show max and average pooling, but realized I had used max pooling for both after doing all the animations 😪

[@rotemlv]

​@@far1din Thanks for the reply. Yeah your logic regarding showing the alternatives makes sense. It's just that from what I read, sigmoid isn't recommended (basically just "use ReLU", since using sigmoid can "kill" the gradient more easily than ReLU). Also (forgot to say this in my comment) - these videos are very informative and easy to follow, kind of a 3b1b vibe, with the animations.

[@far1din]

Thank you. These animations are made with the same library (manim) which 3b1b created! :)

[@TheSquareClasses]

You explained everything in detail with mathematics except the fully connected layers, how they work. Please explain this one in a different video.

[@emadhajaj4245]

Great work, actually it one of the most beautiful videos made in AI.

[@far1din]

Thank you my friend! 😃

[@mennobangma]

So where is the first convolutional Layer based on. Why do these filters work so well on 'analysing' numbers? What kind of edges or shapes do they detect?

[@far1din]

The weight’s were initialized randomly at det beginning and trained for 100 epochs/iterations if I’m not mistaken (this was a couple months ago). The training data is from the mnist dataset, which is a dataset containing only handwritten digits. That’s why this network is capable of detecting handwritten digits. The edges and shapes detected by each of these filters post training can be seen in the activation layers. These are the layers to the right of each filter after the convolutions. For an untrained network that doesn’t have «trained filters» the output will most likely be blurry, have «random pixels» and the model will output «random values». As you train the model (backpropagation) the filters in the model will learn to detect shapes etc. as seen in the video. I should probably make a visualization for the entire process 🤔 I hope this clarified some of your questions!

[@r0cketRacoon]

what happens if I specify the convo layer 2 have only 2 dimensions? the same kernel will be applied for both 2 images? then be added?

[@SolathPrime]

Wow first

[@far1din]

😂🔥

[@hihaoay8042]

Well i have a question, the final result is 7 and 3.6 so the model predict the input number is 7, so how about the others, I mean from 0 to 100 fully connected layer, they can be 0 , 8, 9,10 right ? so what value will they predict? Thank you so much for your video

[@ahmedhesham3125]

good video

[@tejan8427]

How do we know how many layers or filters we need at each layer ? I mean, how can we construct our architecture.

[@BooleanDisorder]

But like, how did we learn to do this? How was the logic of all the layers thought out?

[@윤기좔좔엉덩이]

What are the criteria for setting filters?

[@domahidipeter6092]

Activation 1 layer ishave a dimension of ( 28-5+1=24) 24*24?

[@oculotronicstest2866]

Hi, Can anyone tell me how the weights are assigned in the last fully connected layer. Thanks in advance : )

[@far1din]

Initially it’s «random», but then it get’s trained through the backpropagation!

[@bamboooooooooooo]

does the convolutional layer always have a stride of 1?

[@far1din]

No, it’s something you choose.

[@rubytejackson]

exceptional explanation u did! I have several questions , but first id like to ask is it ok to support u from the thanks button since i dont have any paypal account? thnks warmest regards ruby

[@far1din]

Ofc my friend! Feel free to shoot me a DM on X if you have any questions aswell 💯

[@keremkezer6826]

👏👏👏👏👏👏👏👏👏👏👏👏

[@Aldotronix]

i can't understand how a computer can figure out an image after many convolutions, seems like magic.

[@riturajput9040]

How is weight initialised ?

[@andyh3970]

That is done during training via back propagation Here’s a picture, The answer is 7 now set the weights backwards so the output neuron for 7 turns on.

[@mysteriousXsecret]

7:13 why am I having 4x2 filters?

[@far1din]

Basically, you are free to choose how many filters you want and what size they should be. I choose the filters just so you (viewers) could get a better understanding of the convilution process! Zfj

[@iancoify]

wow ty! -n

[@arpitgaur4310]

this was missing in 3b1b video of CNN

[@nithina5105]

Can you do a video on gradCAM working?

[@far1din]

Unfortunately, I'm not well-versed in the concept of GradCAM. :/

[@nitishaggarwal-i8y]

mast bc

[@way2on]

can you believe this is over complicated. you can literally just do some upscaling. plus make it bi-directional