Backpropagation in Convolutional Neural Networks (CNNs)
In this video we are looking at the backpropagation in a convolutional neural network (CNN). We use a simple CNN with zero padding (padding = 0) and a stride of two (stride = 2).
X: / far1din_
Github: github.com/far1din
Manim code: github.com/far1din/manim#back...
---------- Content ----------
00:00 - Introduction
00:51 - The Forward propagation
02:23 - The BackPropagation
03:31 - (Intuition) Setting up Formula for Partial Derivatives
06:07 - Simplifying Formula for Partial Derivatives
07:05 - Finding Similarities
08:55 - Putting it All together
---------- Contributions ----------
Background music: pixabay.com/users/balancebay-...
#computervision #convolutionalneuralnetwork #ai #neuralnetwork #deeplearning #neuralnetworksformachinelearning #neuralnetworksexplained #neuralnetworkstutorial #neuralnetworksdemystified #computervisionandai #backpropagation
great video, but i don't understand how we can find the value of the dL/dzi terms. At 7:20 you make it seem like dL/dzi = zi, is that correct?
No, they come from the loss function. I explain this at 4:17. It might be a bit unclear so I’ll highly reccomend you watch the video from 3blue1brown: kzhead.info/sun/p62eeLCmoaVriHA/bejne.htmlsi=Z6asTm87XWcW1bVn 😃
I'm with @louissimion, you show how dL/dw1 is related to dz1/dw1+... (etc), but you never show/expain where dL/dz1 (etc) comes from. Poof - miracle occurs here. Having a numerical example would help a lot. This "theory/symbology" only post is therefore incomplete/useless from a learing/understanding standpoint.
Really intuitive and great animations.
Fantastic explanation!! Very clear and detailed, thumbs up!
great job. this explanation is really intuitive
Great Explanation, helped me understand the background working
great stuff man, crystal clear!
what i was looking for. well explained
This channel is a hidden gem. Thank you for your content
excellent. the exact video i was looking for.
Couldn’t explain it better myself … absolutely amazing and comprehensible presentation!
This is a topic which is rarely explained online, but it was very clearly explained here. Well done.
This was really helpful....Thank you so much for the vizualization...Keep up the good work...Looking forward to your future uploads.
great explanation, clear direct and understandable, sub!
really clear explanation and good pacing. I felt I understood the math behind back propagation for the first time after watching this video!
Nicely put, thank you so much.
great video, underrated channel , please keep it up with CNN videos!
great explanation
the animations were super useful, thanks!
Thanks for sharing!
really beautiful, thanks.
What a masterpiece.
Very well explanation, I search many videos but no body explained regarding change in filter's weight. Thank you so much for this animated simple explanation.
Best explanation
Great example thanks a lot
I have seen few videos before, this one is by far the best one. It breaks down each concept and answers all the questions that comes in the mind. The progression, the explanation is best
Thank you! 🔥
Well done.
Why is this channel so underrated? You deserve more subscribers and views.
Perhaps developers use ad blockers, and as a result, KZhead needs to ensure revenue by not promoting these types of videos (that's my opinion)
please continue your videos !!
your channel is a Hidden Gem..My suggestion is to start a discord and get some crowd functing and one on ones for people who want to learn from you..youa re gifted in teaching.
amazing video thanks!
Great explanation with cool visual. Thanks a lot.
Thank you my friend 😃
Amazing! I was looking for some material like this a long time ago and only found it here, beautiful :D
Thank you my brother 🔥
Masterpiece 💕💕
Thank you so much!!! This video is so so so well done!
Thank you. Hope you got some value out of this! 💯
Great explanation and visualization
Thank you my friend 🔥🚀
thanku you so much for this
Please do not stop making these videos!!!
I won’t let you down Joker 🔥🤝
The equation at 6:00 ends up with ∂L/∂w(i) = 4 * ∂L/∂w(i) if we cancel out ∂z. It is Multivariable chain rule so the correct function is: ∂L/∂w(i) = ∂z1/∂w(i) * dL/dz1 + ∂z2/∂w(i) * dL/dz2 + ∂z3/∂w(i) * dL/dz3 + ∂z4/∂w(i) * dL/dz4. So we can't do cancelation.
You are a great example of fluidity of thought and words..great explanation
Thank you my friend. Hope you got some value! :)
@@far1din619 sure did
Amazing
fab video! help me a lot
Glad to hear that you got some value out of this video! :D
Well explained now I need to code it my self
Haha, that’s the hard part
@@far1din619 I think I came up with a solution Here def backward(self, output_gradient, learning_rate): kernels_gradient = np.zeros(self.kernels_shape) input_gradient = np.zeros(self.input_shape) for i in range(self.depth): for j in range(self.input_depth): kernels_gradient[i, j] = convolve2d(self.input[j], output_gradient[i], "valid") input_gradient[j] += convolve2d(output_gradient[i], self.kernels[i, j], "same") self.kernels -= learning_rate * kernels_gradient self.biases -= learning_rate * output_gradient return input_gradient First i initialized the kernel gradient as an array of zeros with the kernel shape then I iterated through the depth of the kernels the the depth of the input then for each gradient withe respect to the kernel I did the same to compute the input gradients Your vid helped me understand the backward method better So I have to say thank you sooo much for it
@@far1din619 I'll document the solution and but it here when I do please pin the comment
@@SolathPrime That’s great my friend. Will pin 💯
Thanks.
Great video!! Your explanation is the best I have found. Could you please tell me what software you use for the animations ?
I use manim 😃 www.manim.community
thx
I've had no trouble learning about the 'vanilla' neural networks. Although your videos are great, I can't seem to find resources that delve a little deeper into the explanations of how CNNs work. Are there any resources you would recommend ?
What about the weights of the fully connected layer
+1 sub, excellent video
Thank you! 😃
perfect, one suggestion make videos a little longer 20-30 is a good number
Haha, most people don't like these kind of videos too long. Average watchtime for this video is about 3minutes :P
@@far1din619oh shii! 3 minutes, that was very unexpected, maybe it's because people revisit the video to revise specific topic.
Must be 💯
Great explanation. Can you please tell which tool do you use for making these videos.
Thank you my friend! I use manim 😃 www.manim.community
What is the loss function here, and how are the values in the flattened z matrix used to compute yhat ?
1:15 why do you iterate in steps of 2? If you iterated by 1 then you could generate a 3x3 layer image. Is that just to save on computation time/complexity or is there something other reason for it?
The reason why I used a stride of two (iterations in steps of two) in this video is partially random and partially because I wanted to highlight that the stride when performing backpropagation should be the same as when performing the forward propagation. In most learning materials I have seen, they usually use a stride of one, hence a stride of one for the backpropagation. This could lead to confusion when operating with larger strides. The stride could technically be whatever you like (as long as you keep it within the dimensions of the image/matrix). I could have chosen another number for the stride as you suggested. In that case, with a stride of one, the output would be a 3 x 3 matrix/image. Some will say that a shorter stride will encapsulate more information than a larger one, but this becomes “less true” as the size of the kernel increases. As far as I know there are no “rules” for when to use larger strides and not. Please let me know if this notion has changed as everything changes so quickly in this field! 🙂
@@far1din619 I never considered how stride length could change depending on kernel size. I guess that makes sense, the larger kernel could cover the same data as a small kernel, just in fewer steps/iterations. I also figured you intentionally generated a 2x2 image since that’s a lot simpler than a 3x3 and this an educational video. Thanks for the feedback, that was really insightful!
5:24 does this just mean we divide z1 by w1 and ultiply by L divided by z1 and do that for all z'S to get the partial derivative of L in respect to w1?
It’s not that simple. Doing the actual calculations is a bit more tricky. Given no activation function, Z1 = w1*pixel1 + w2*pixel2 + w3*pixel3… you now have to take the derivative of this with respect to w1, then y = z1*w21 + z2*w22… take the derivative of y with respect to z1 etc. The calculus can be a bit too heavy for a comment like this. I’ll highly reccomend you watch the video by 3blue1brown: kzhead.info/sun/p62eeLCmoaVriHA/bejne.htmlsi=Z6asTm87XWcW1bVn 😃
Hello well explained. I need your presentation
Just download it 😂
dL/dzi = ??
I explain the term at 4:17. It might be a bit unclear so I’ll highly reccomend you watch the video from 3blue1brown: kzhead.info/sun/p62eeLCmoaVriHA/bejne.htmlsi=Z6asTm87XWcW1bVn 😃
You have nices videos, that helped me better understand the concept of CNN. But, from this video, it is not really obvious that matrix dL/dw - is convolution of image matrix and dL/dz matrix, as showed here kzhead.info/sun/g9Jwgq9vq6Gcg58/bejne.html. The stride of two is also a little bit confusing
Thank you for the comment! I believe he is doing the exact same thing (?) I chose to have a stride of two in order to highlight that the stride should be similar to the stride used during the forward propagation. Most examples stick with a stride of one. I now realize it might have caused some confusion :p
w^* is an abuse of math notation, but it's convenient.
I think it's spelled "Convolution"
Haha thank you! 🚀