What backpropagation actually is?

I have a conceptual question due to terminology that bothers me. Is backpropagation algorithm a neural network training algorithm or is it just a recursive algorithm in order to calculate a Jacobian for a neural network? Then this Jacobian will be used as part of the main training algorithm e.g Steepest Descent?

Hence is it a training algorithm or a numerical way to calculate a Jacobian matrix (partial derivatives of neural network outputs respective to network parameters)?

No, I wouldn't consider backprop a training algorithm. Backpropagation is just a way to find the derivative of the loss function with respect to the inputs by using the chain rule. Computing a derivative doesn't train anything.

What you do with this derivative in order to minimize the loss function is the training part.

EDIT: I think it will depend on who you ask. Take for example, this PyTorch tutorial. They say that "Backward propagation: In backprop, the NN adjusts its parameters proportionate to the error in its guess. It does this by traversing backwards from the output, collecting the derivatives of the error with respect to the parameters of the functions (gradients), and optimizing the parameters using gradient descent."

I.e. the two steps

loss.backward()
optim.step()


together are what they call backpropagation. This is what I'd call the more engineering view point and I believe is a semantic shift away from what I'd argue (see comments!) is actually backprop and that's just the loss.backward() step.

The semantic drift of backprop meaning calculating the derivatives together with optimization makes sense in this context. Why would you call loss.backward() and then not call optim.step()? But, originally (and technically, the best kind of correct) backprop refers to just the computation of the derivatives and I'll think you'll find that terminology more in math/theory contexts instead of the programming/engineering contexts.

• it is not an opinion, it is an accepted fact; backpropagation is an algorithm used to train the model, i.e. learn the model weights, etc... Look for well known references :) Sep 8 '21 at 15:09
• You are not correct. Backprop merely computes the gradient. Check references for yourself. For example Goodfellow, Bengio, Courville "Backprop allows information from the cost to then flow backward through the network in order to computer the gradient" and "The term back-propagation is often misunderstood as meaning the whole learning algorithm for multi layer NNs. Actually, backprop refers *only* to the method for computing the gradient, while another algorithm, such as SGD is used to perform learning using this gradient." Sep 8 '21 at 15:36
• Even from the Chollet book you refer to, "Backpropagation is a way to use the derivative of simple operations (such as addition, relu, or tensor product) to easily compute the gradient of arbitrarily complex combinations of these atomic operations." Sep 8 '21 at 15:40
• If you click on the backprop link from the tensorflow site it leads to the wiki which agrees with me. "backpropagation computes the gradient of the loss function with respect to the weights of the network for a single input–output example". Continuing, "and does so efficiently, unlike a naive direct computation of the gradient with respect to each weight individually." It's just an efficient way to compute derivatives. Yes, things are trained via (aka using) backprop. As seen in the next sentence of the wiki "This efficiency makes it feasible to use gradient methods for training...". Sep 8 '21 at 16:02
• You can say "I built a deck with a hammer" and it makes sense, but a hammer just drives nails. It doesn't build decks directly. Where you drive the nails doesn't concern the hammer. Backprop computes derivatives. If you use gradient methods to train a network doesn't matter to backprop. Backprop just computes the gradients. Sep 8 '21 at 16:04

Backpropagation algorithm is the way the neural network weights are optimized (learned), i.e., what the optimizer uses for this purpose, so yes it can be considered the training algorithm.

In backpropagation, you do not need to explicitely calculate the Jacobian matrix (see this source of info for more detail), but you calculate the first derivatives of your loss function with respect to its weights via the chain rule (from calculus). This chain rule lets us find the gradient of more complex functions by splitting it into simpler derivatives.
Then, by applying the gradient descent strategy (following the steepest descent, i.e., adding the negative derivative at hand for each trainable weight), the weights are adjusted in each iteration until certain values generate a loss which can be accepted or until a certain number of rounds is reached.

As a simple example of how each derivative would be used to update each weight, we can think of linear regression applying gradient descent:

As clearly pointed out in the fantastic book Deep learning with python by François Chollet: "Backpropagation starts with the final loss value and works backward from the top layers to the bottom layers, computing the contribution that each parameter had in the loss value".

• I should point out that the quote you reference at the end reworded says "Backprop starts with C(y) (y are the outputs) and works backwards to compute \frac{\partial C}{\partial w} and \frac{partial C}{partial b}" I.e. it just computes gradients. And the linked pdf just computes \frac{\partial L}{\partial x} and makes no mention of what to do with the derivative. I.e. read the first sentence of the second paragraph of chapter 2. "we assume that we are given ∂L/∂y and our goal is to compute ∂L/∂x and ∂L/ ∂w". That's all backprop does. Sep 8 '21 at 15:56

I interpret the things in away that the terminology has been a bit twisted over several sources. Some refer to the backpropagation as a training algorithm and some just as an efficient algorithm to compute the partial derivatives (i.e the Jacobian).

Hence, I believe the original meaning for backpropagation algorithm was to calculate the Jacobian for the network. Then any training algorithm could utilize the Jacobian. It seems that most often the chosen training algorithm is steepest descent.

• This answer would be better if it included some references; at the moment it just sounds like personal opinion. Either way, I feel it would have been better as part of your question. Sep 13 '21 at 18:25