Backpropagation in CNN - PART 2
Vložit
- čas přidán 27. 07. 2024
- Backpropagation in CNN is one of the very difficult concept to understand. And I have seen very few people actually producing content on this topic.
So here in this video, we will understand Backpropagation in CNN properly. This is part 2 of this tutorial, and in this is we will look at Backpropagation for entire Convolutional Neural Network. In part 1, we already saw the backpropagation for convolutional operation. You can find its link down here in the description box.
All the frameworks used for Deep Learning automatically implement Backpropagation for CNN. But as we humans are curious, we want to know how it works and not let it be implemented automatically.
So buckle up! And let's understand Backpropagation in CNN.
➖➖➖➖➖➖➖➖➖➖➖➖➖➖➖
📕 PDF notes for this video: bit.ly/BackPropCNNP2
➖➖➖➖➖➖➖➖➖➖➖➖➖➖➖
Timestamp:
0:00 Intro
1:52 Forward Propagation Equations
3:23 What to obtain
4:52 Layer 3 Backpropagation
7:17 Layer 2 Backpropagation
13:41 Layer 1 Backpropagation
17:47 Summary
➖➖➖➖➖➖➖➖➖➖➖➖➖➖➖
📕 Backpropagation in CNN - Part 1: • Backpropagation in CNN...
📕 Neural Network (ANN) Playlist: • How Neural Networks wo...
➖➖➖➖➖➖➖➖➖➖➖➖➖➖➖
Follow my entire playlist on Convolutional Neural Network (CNN) :
📕 CNN Playlist: www.youtube.com/watch?v=E5Z7F...
At the end of some videos, you will also find quizzes 📑 that can help you to understand the concept and retain your learning.
➖➖➖➖➖➖➖➖➖➖➖➖➖➖➖
✔ Complete Logistic Regression Playlist: • Logistic Regression Ma...
✔ Complete Linear Regression Playlist: • What is Linear Regress...
➖➖➖➖➖➖➖➖➖➖➖➖➖➖➖
If you want to ride on the Lane of Machine Learning, then Subscribe ▶ to my channel here: / @codinglane
Dude! People in universities get paid SO MUCH to explain this stuff, and you're providing it online, and even in a better form! Lifesaver for high-schoolers like me. I did not find any better CNN backprop video, and I found you. What you are doing is CRYSTAL clear and I ended up checking out the entire series, and I'm eager to watch more. We need more people like you, damn.
I am truly glad to hear this. Thanks a lot!! ❤️
Man, this is best series i have ever watched. Thanks a lot :)
My god I have to watch this two part derivation series god knows how many times. Can only imagine how tedious it was for you to expunge this information which honestly is rare in most CNN tutorials
Thank you so much! Yes, it took a lot of effort. Glad it was worth it and you found it valuable.
This is the only content in online Tutorials explaining CNN training, specifically Convolutional Filter updation process.
Very much impressive with in detail explanation.
All the best. Keep do it . Let's Grow Together..
Thank you so much for your words. Highly appreciate it!! 🙂
Literally the best CNN tutorial out here on CZcams! Thanks a lot!
This is remarkable explanation of Backprop in CNN. But there is a slight mistake in the slide in the chain rule of dZ[1] which is equals dL/dC[1] * dC[1]/dZ[1] not dZ[2].
Actually, this is the best explanation that I looked for in the entire Internet. thx
Thank you for saving me as I have my AI course final in 2 days and I am finally undertanding everything in a much clearer and cohesive way
You are amazing, and your content for convolution neural network is mind blowing, I am really lucky I found your channel, Your work is appreciated, Thank you so much !
Very glad to hear this! Thanks!!
hey man, you are a genius!!! even books don't have as good explanation as yours. God bless you! all the best
What a nice video man. You gave such a clear explanation. Thank you!!
Thank you very much for your work!!!!! Your videos are so nice and detailed, that i build from scratch my own CNN.
The backpropagation concept has been explained so clearly. It helps me a lot. Could you please make a video on complex-valued backpropagation on CNN?
This really helps i got my semester exam and im gonna recommend this to my friends as well
thanks so much! this couple of videos helped me out a lot. cheers :)
Amazing explaination !!!!! Gotta give u a thumb up!
Thank you bro, it is really a lot of effort.
love this video!
Great to see the effort put into
Great job.
you are a life saver
Haha, glad I could help! 🙂
actually excellent
Great video!
You are doing great work make complete a.i roadmap video if you can please
thankyou soo much for your efforts. Thus...
Man, can you PLs make a session on Depthwise Conv ?
Superb Explanation !!!!!
Thank you!
Thank you so much..😄 I learned a lot from your video series.
Thanks again ❤
You’re welcome 😇
Thansk , appreciate bro
NIcely explained
thank you sir
Its 👏 amazing
Thank you:)
Top video
Hi @Coding Lane, I dont get the part at 7:05 dL/df. Can you explain where did f2'(Z2) go?
Hi bro, everything was great. I learnt a lot from it. Can you explain the process in conv layer with an example just as you did for max pool layer. Thanks in advance
Thanks for the suggestion… will try to cover that in some other video
Hi, I had a question. Do we calculate the gradient with respect to lower layer with the updated kernel (using the gradient with respect to the weights) or using the existing kernel values for the layer?
Can u please suggest some reading material for such mathematics?
Hey Lane! I went back trough the video, and I can't wrap my head around the matrixes at 12:21. Is it not supposed to be dC2/Z2 instead of dC2/dZ2? Or, is the matrix with the shown values supposed to be labelled dZ2? I acknowledge that there might be a misunderstanding from my side, however, could you help me clarify this? I know you are checking the comments, which makes you even a better teacher! Thank you in advance!
Hi, it will be dC2/dZ2 only. Its the derivative chain multiplication method.
One thing I don't understand is, for example:
Convolution layer input image has dimensions (batch=1, width=5, height=5, channels=3),
Convolution layer has 2 kernels each (width=3, height=3, channels=3), we use valid padding and stride=2,
Convolution layer output will have dimensions (batch=1, width=3, height=3, channels=2),
When the gradients come back to the convolution layer, they will have the same dimensions as the output.
Then how is it possible to convolve a (1, 5, 5, 3) image with the (1, 3, 3, 2) gradients when the image has channels=3 but gradients have channels=2 (where 2 is the number of kernels used)?
What is the name of the loss function here?
Thank you for your great explanation 👍
Do you can give me some sources for where you got that knowledge from?
I mostly learned from Andrew Ng’s coursera courses. You can checkout DeepLearning.ai courses… specifically the ones given by Andrew Ng
hello.. at 12:53, for calculating dK, we do the convolution process... i have one doubt... in my case, at the 1st conv layer, i use 2kernels for 1 image, so it gives me 2 outputs.... now when it reaches 2nd conv layer, i use 8 kernels (as 2 inputs, 1 for each input).. when we do backprop, if i have to convolve P[1] with dZ[2] , how can i do it.. coz i have 2 inputs (P[1]) and 4 matrices in dZ[2]?? also if this is possible, then how will i update the kernel/weight values as i used 8 but now im using only 4 for convolving... pls reply...
What I can conclude from the video is we need access to the outputs of all the intermediate layers (not just the final output) in backpropogation. Is this correct?
thankyou bro, very helpful. i d like to ask, in the time 8:25 you say that to get dP, just reshape df to the shape P. but what if the flatten layer is repleced by global average pooling?
Hi Mark.. I explained calculating derivative for average pooling layer as well... so similar computation will follow. But note that flatten layer is essential in CNN, we must have flatten layer before fully connected layers.
Hi, Thank you for these vidéos my questions are what is the relationship between weights and kernels in this backpropagation
Hi… Kernels/filters have weights as its parameters
@@CodingLane Yes I know, but I can't figure out how the filters is going to be updated based on the connection weights, what I have known that the filter components are the connection weights right ?
👍👍👍👍👍👍👍👍
Appreciate it. Very good content. I have a request, can you please share pdf of this part? as you have shared pdf of previous part but this video's pdf is not in the description. so please share. Thanks!!!
Thanks for bringing this to my attention... I will share the pdf soon
I learn a lot cause i have to create own cnn from scratch
I am glad that these videos helped you 😇.
What do you mean by dot product multiplication at 12:32, can you explain by small example
Sure... here is an example.
The dot product multiplication of
[[1,2],
[3,4]]
*
[[2,3],
[4,5]]
=
[[2,6],
[12,20]]
I guess, it's better to say element-wise multiplication than dot product multiplication.
Thank you Coding Lane for your great work 👍.
No word for you😢
China is No 1 economy in the essence of political stability, economic purchase power, AI guided cheap quality industrial products, technologically, educationally, economically, by worldwide cooperation, militarily.
India is No 2,
Russia is No 3.
All wishes for BRICS and allies.
Hi… greetings!!
Oh boy...May God save us