L6.2 Understanding Automatic Differentiation via Computation Graphs

As previously mentioned, PyTorch can compute gradients automatically for us. In order to do that, it tracks computations via a computation graph, and then when it is time to compute the gradient, it moves backward along the computation graph. Actually, computations graphs are also a helpful concept for learning how differentiation (computing partial derivatives and gradients) work, which is what we are doing in this video.
Slides: sebastianraschka.com/pdf/lect...
-------
This video is part of my Introduction of Deep Learning course.
Next video: • L6.3 Automatic Differe...
The complete playlist: • Intro to Deep Learning...
A handy overview page with links to the materials: sebastianraschka.com/blog/202...
-------
If you want to be notified about future videos, please consider subscribing to my channel: / sebastianraschka

[@karimelkabbaj]

Thank you very much for this simplified explanation, i've been struggling to understand it until i found this master piece.

[@SebastianRaschka]

Nice, glad to hear that this was useful!!

[@manuelkarner8746]

thaaaank you, finaly I understand this perfectly (& can know repeat it for myself)
explaining backpropagation my lovely proffs always said "then this is just the chainrule" & skipped any explanation
for calculating (complicated) toy examples I knew the chainrule, but in the backprop context it was just to confusing
________Anway, got a question: at 12:23 you said tehcnicaly canceling the delta terms is not allowed -> could you elaborate on the math/why or point me to some ressourece explaining this ?
Intuitively I always thought canceling delta´s is strange/unformal but I dont found out how this delta notation stuff fits into "normal" math notation :)

[@SebastianRaschka]

Nice, I am really glad to hear that! And yes, I totally agree. When I first learned it, it was also very confusing at first because the prof tried to brush it aside ("it's just calculus and the chain rule") just like you described!

[@nak6608]

Love your textbooks and your videos. Thank you!

[@mahmoodmohajer1677]

thanks for pulling up this video.

[@user-kw4kp7eq9m]

Thank you very much!

[@Gzzzzzz111]

YOU ARE GOATED!

[@736939]

17:27 In the formula on the top-left (as I understood) there is no sum, but stacking (or concatenating), then why should we add the results in different paths during the backward chain computation? Is it always work like this - just produce the sum in the chain when there is a concatenating????

[@SebastianRaschka]

Sorry if this was misleading. In the upper left corner, this was more like a function notation to highlight the function arguments. Like if you have a function L that computes x^2 + y^2, then it's basically like writing L(x, y) = x^2 + y^2. There is no concatenation. With the square brackets I meant to show that sigma_3 contains also function arguments. I just used square brackets (instead of round brackets) so it is easier to read, but now I can see how this can be confusing.

[@736939]

@@SebastianRaschka Thank you very much.

[@Epistemophilos]

At 11:27 it gets confusing because you switch the terms around. Otherwise, very nice video.