1
$\begingroup$

I would appreciate some help on the following problem: I'm taking Hinton's coursera class on Neural Nets and I'm not sure I understand the step highlighted in the picture (see below).

Image

Background:

  • i is hidden layer
  • j is the top layer
  • Neurons use logistic regression as their activation function

What I understand:
The chain rule allows you to "break down" the partial derivative and introduce a term that is helpful for your calculation:

formula

What I don't understand:
Where does the formula come from? In other words, what's the proof that you can break down the left term into the sum of 3 components of the top layer (in this case).

Thanks for your help

Link to class: https://www.coursera.org/learn/neural-networks/lecture/gcNo6/the-backpropagation-algorithm-12-min

$\endgroup$
  • $\begingroup$ It's the chain rule for partial derivatives. Essentially, the derivative in the direction of a given vector is the sum of the derivatives of the components of the vector, weighted appropriately. $\endgroup$ – ConMan Jan 30 '17 at 3:05
  • $\begingroup$ Hi ConMan. Thanks for your comment. I understand the chain rule. Can you explain the steps in between ? The chain rule only allows you to "slide in" one zj. Why the sum over j ? $\endgroup$ – Guillaume G Jan 30 '17 at 3:09
  • $\begingroup$ It's a specific version of the chain rule for functions of multiple variables (or equivalently, for functions of vectors). The following link from Khan Academy shows a demonstration (although it looks at the total derivative of the main function, and it's only applied to a 2-variable case): khanacademy.org/math/multivariable-calculus/… $\endgroup$ – ConMan Jan 30 '17 at 3:13
  • 1
    $\begingroup$ @ConMan It is called the total derivative : if $g(t) = f(x(t),y(t))$ then $\frac{\partial g}{\partial t} = \frac{\partial f}{\partial x}\frac{\partial x}{\partial t}+\frac{\partial f}{\partial y}\frac{\partial y}{\partial t}$ $\endgroup$ – reuns Jan 30 '17 at 3:28
  • 1
    $\begingroup$ @GuillaumeG - I suggest you read up on the Jacobian matrix and the chain rule in higher dimensions. Essentially, the ordinary derivative applies when you've got a single underlying variable that you're differentiating with respect to, and the partial derivatives apply when you're only differentiating with respect to one variable out of many. The Wikipedia article on the chain rule, particularly the section on higher dimensions, explains it a bit better: en.wikipedia.org/wiki/Chain_rule#Higher_dimensions $\endgroup$ – ConMan Jan 30 '17 at 4:19
0
$\begingroup$

This follows from the multivariate chain rule. Let's assume that $E$ depends on $z=(z_1,\ldots,z_n)$, but each $z_k$ depends on $y=(y_1,\ldots,y_m)$, i.e. $$ E(z(y))=E(z_1(y),\ldots,z_n(y)). $$ We are interested in how the loss objective $E$ depends on some value in an earlier layer, say the output of the $i$th node, denoted $y_i$, in layer $\xi$, but that output affects $E$ only by its influence on the next layer (i.e., $\xi+1$), whose values we denote $z$. Recall that in a fully connected network, each layer depends on all the values of the previous layer, so each $z_i$ depends on each $y_k$. But we only care about $y_i$ for now, meaning we can ignore the other $y_k$, because we are interested in how perturbing $y_i$ affects $E$, and $y_i$ does not affect $E$ via the other $y_k$.

But $y_i$ does affect $E$ through $z$ (and only through $z$ actually). And $y_i$ affects every $z_k$. So we need to take the $\partial E / \partial z_k$ into account, as well as the $\partial z_k / \partial y_i$. The way to combine these into a single number is specified by the multivariate chain rule: $$ \frac{\partial E}{\partial y_i} = \sum_j \frac{\partial E}{\partial z_j} \frac{\partial z_j}{\partial y_i}. $$ I'll add links to more formal derivations below. But the rough intuition is this: $E$ depends on $y_i$ through its effect on each $z_k$ (captured by $\partial z_k / \partial y_i$), so ${\partial E}/{\partial y_i}$ should be a sum of these contributions (if we think of derivatives as a measure of local linear change, we can imagine adding the perturbation vectors induced by tweaking together). But not all the contributions to the network are equally important: if a $z_j$ doesn't affect $E$ much (i.e.,${\partial E}/{\partial z_j}$ is small), then the contribution of $y_i$ through $z_j$ should be down-weighted. So we do a weighted sum over $\partial z_k / \partial y_i$, with weights given by ${\partial E}/{\partial z_j}$. Notice that when $n=1$ we recover the classical chain rule. For more details see the links below.


Other questions on the origin of the multivariate chain rule are here, here, and here.

A related question on the role of the chain rule sum in artificial neural networks is here.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.