This is the back-propogation rule for the output layer of a multi-layer network:

$$W_{jk} := W_{jk} - C \dfrac{\delta E}{\delta W_{jk}}$$

What does this rule do in the more ambiguous cases such as:

(1) The output of a hidden node is near the middle of a sigmoid curve?

(2) The graph of error with respect to weight is near a maximum or minimum?

  • 1
    $\begingroup$ Please format any mathematics using MathJax, rather than using images. $\endgroup$ – The Pointer Jan 1 at 15:13

I assume you are considering a network where the activation function of the last layer is a sigmoid, so the output of your network is $$\tilde{y}=\sigma(W^{L}\cdot f(X, W^1, \dots, W^{L-1})),$$ where $X$ is the input vector, and $f$ is obtained by feeding the input to the network up to the layer $L-1$. Let's also call $Z:= W^{L}\cdot f(X, W^1, \dots, W^{L-1})$.

The error term is computed as $$E(y, \tilde{y})=E(y, \sigma(Z)),$$ where $y$ is the actual output. Let's get the derivative of the error with respect to the output of the last node (the input of the sigmoid) $$\frac{\partial E}{z_i}=\frac{\partial E}{\partial \tilde{y}}\frac{\partial\tilde{y}}{\partial z_i}=\frac{\partial E}{\partial \tilde{y}}\frac{\partial\sigma}{\partial z_i}.$$ The update rule is $$z_i =z_i - C\frac{\partial E}{z_i}= z_i - C\frac{\partial E}{\partial \tilde{y}}\frac{\partial\sigma}{\partial z_i}.$$ Now we can analyse your questions.

  1. To be close to the middle of the sigmoid means that $z_i$ is close to $0$; moreover the derivative $\frac{\partial\sigma}{\partial z_i}$ reaches its maximum value when it is evaluated in $0$. This means that the term $\frac{\partial\sigma}{\partial z_i}$ gets "large" as $z_i$ approaches the center of the sigmoid, contributing more to the update of the weight. Of course it is difficult to say what happens in general, as the term $\frac{\partial E}{\partial \tilde{y}}$ is also in the expression and it is possible that this term gets really small (or big) for $z_i$ close to $0$. You just know that the term $\frac{\partial\sigma}{\partial z_i}$ is trying to push $z_i$ further away from $0$, so the idea should be that the closer to the center the larger the update.
  2. To be close to an extremum point of the loss means that the derivative of the loss with respect to $\tilde{y}$ is close to $0$. Since $\frac{\partial E}{\partial \tilde{y}}$ is backpropagated in a multiplicative fashion, the rule of thumb is that the closer you are to an extremum the smaller the updates get. Though, as above, is kind of difficult to say what will happen when updating a general node, as some of the term in the multiplication can be very large, making the update large even if $\frac{\partial E}{\partial \tilde{y}}$ is small.

For instance, what happens if you are close to the center of the sigmoid but also close to an extremum of the loss? You will have a multiplication of $2$ terms, one trying to make the update small and the other trying to make the update large, and what matter are the orders of magnitude involved.
In conclusion, the rules of thumb are as in the points 1. and 2., but they are no guarantee that you won't find any special cases.


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.