What is the difference between SGD and back-propagation?

Back-propagation is just a method for calculating multi-variable derivatives of your model, whereas SGD is the method of locating the minimum of your loss/cost function.


Stochastic gradient descent (SGD) is an optimization method used e.g. to minimize a loss function.

In the SGD, you use 1 example, at each iteration, to update the weights of your model, depending on the error due to this example, instead of using the average of the errors of all examples (as in "simple" gradient descent), at each iteration. To do so, SGD needs to compute the "gradient of your model".

Backpropagation is an efficient technique to compute this "gradient" that SGD uses.


Backpropagation is an efficient method of computing gradients in directed graphs of computations, such as neural networks. This is not a learning method, but rather a nice computational trick which is often used in learning methods. This is actually a simple implementation of chain rule of derivatives, which simply gives you the ability to compute all required partial derivatives in linear time in terms of the graph size (while naive gradient computations would scale exponentially with depth).

SGD is one of many optimization methods, namely first order optimizer, meaning, that it is based on analysis of the gradient of the objective. Consequently, in terms of neural networks it is often applied together with backprop to make efficient updates. You could also apply SGD to gradients obtained in a different way (from sampling, numerical approximators etc.). Symmetrically you can use other optimization techniques with backprop as well, everything that can use gradient/jacobian.

This common misconception comes from the fact, that for simplicity people sometimes say "trained with backprop", what actually means (if they do not specify optimizer) "trained with SGD using backprop as a gradient computing technique". Also, in old textbooks you can find things like "delta rule" and other a bit confusing terms, which describe exactly the same thing (as neural network community was for a long time a bit independent from general optimization community).

Thus you have two layers of abstraction:

  • gradient computation - where backprop comes to play
  • optimization level - where techniques like SGD, Adam, Rprop, BFGS etc. come into play, which (if they are first order or higher) use gradient computed above