I am a little confused when trying to find the gradient for the multiclass hinge loss:

$$l(y) = \max( 0, 1 + \underset{r \neq y_i}{ \max } W_r \cdot x_i - W_{y_i} \cdot x_i)$$

Where $W^{k \times n}$ is the matrix holding in each row the corresponding classifier of each class.

Unless my math is wrong, the gradient of the function is: \begin{equation} \frac{\partial l}{\partial w} = \begin{cases} 0, & W_{y_i}\cdot x_i > 1 + W_r \cdot x_i \\ 0 + 0 - x_i, & \text{otherwise} \end{cases} \end{equation}

Is this ok?

So if I would like to find the $W$ which minimizes the function using the stochastic gradient descend I would need to do: \begin{equation} W_y^{(t+1)} = W_y^{(t)} - \eta x_i \end{equation}

with $\eta$ the learning rate.

Is this a valid procedure to optimize the $l(y)$ function?

  • $\begingroup$ What is $w$? And since in definition $l$ depends on $y$, shouldn't then $\frac{\partial l}{\partial w}=0$? $\endgroup$ – mpiktas Jun 2 '15 at 6:20
  • $\begingroup$ $\omega$ is the same as $W_{y_i}$. $l$ is also dependent on $W$ and $x_i$. I wrote it that way because that's how I saw it in another places. $\endgroup$ – Eliel Van Hojman Jun 2 '15 at 6:27
  • $\begingroup$ And what is $W_{y_i}$? If you define all the quantities in the formula you might get answers from people who are not aware what is a multiclass hinge loss. Derivatives is a part of calculus, and much more people know calculus. $\endgroup$ – mpiktas Jun 2 '15 at 6:31
  • $\begingroup$ $W_{y_i}$ is the row y of the matrix $W$. I think you are right and I might be confusing some concepts. What I would like to know is if there is a straight-forward solution to the multiclass problem or if the solution is to treat them in a one vs. all style. $\endgroup$ – Eliel Van Hojman Jun 2 '15 at 6:41

Let's use the example of the SVM loss function for a single datapoint:

$L_i = \sum_{j\neq y_i} \left[ \max(0, w_j^Tx_i - w_{y_i}^Tx_i + \Delta) \right]$

Where $\Delta$ is the desired margin.

We can differentiate the function with respect to the weights. For example, taking the gradient with respect to $w_{yi}$ we obtain:

$\nabla_{w_{y_i}} L_i = - \left( \sum_{j\neq y_i} \mathbb{1}(w_j^Tx_i - w_{y_i}^Tx_i + \Delta > 0) \right) x_i$

Where 1 is the indicator function that is one if the condition inside is true or zero otherwise. While the expression may look scary when it is written out, when you're implementing this in code you'd simply count the number of classes that didn't meet the desired margin (and hence contributed to the loss function) and then the data vector $x_i$ scaled by this number is the gradient. Notice that this is the gradient only with respect to the row of $W$ that corresponds to the correct class. For the other rows where $j≠{{y}_{i}}$ the gradient is:

$\nabla_{w_j} L_i = \mathbb{1}(w_j^Tx_i - w_{y_i}^Tx_i + \Delta > 0) x_i$

Once you derive the expression for the gradient it is straight-forward to implement the expressions and use them to perform the gradient update.

Taken from Stanford CS231N optimization notes posted on github.

| cite | improve this answer | |
  • 1
    $\begingroup$ Can you send me your code at ckm.jnu@gmail.com? $\endgroup$ – CKM Sep 22 '16 at 8:32
  • $\begingroup$ "We can differentiate the function with respect to the weights" - Can you please explain how you developed that formula from the Loss function? as in what are the steps / derivation rules used... I can't figure this one out. And... why do you do it w.r.t Wj AND Wyi? Thanks $\endgroup$ – Uri Abramson Mar 29 '17 at 8:16

First of all, note that multi-class hinge loss function is a function of $W_r$. \begin{equation} l(W_r) = \max( 0, 1 + \underset{r \neq y_i}{ \max } W_r \cdot x_i - W_{y_i} \cdot x_i) \end{equation} Next, max function is non-differentiable at $0$. So, we need to calculate the subgradient of it. \begin{equation} \frac{\partial l(W_r)}{\partial W_r} = \begin{cases} \{0\}, & W_{y_i}\cdot x_i > 1 + \underset{r \neq y_i}{ \max } W_r \cdot x_i \\ \{x_i\}, & W_{y_i}\cdot x_i < 1 + \underset{r \neq y_i}{ \max } W_r \cdot x_i\\ \{\alpha x_i\}, & \alpha \in [0,1], W_{y_i}\cdot x_i = 1 + \underset{r \neq y_i}{ \max } W_r \cdot x_i \end{cases} \end{equation} In the second case, $W_{y_i}$ is independent of $W_r$. Above definition of subgradient of multi-class hinge loss is similar to subgradient of binary class hinge loss.

| cite | improve this answer | |
  • $\begingroup$ Could you please clarify what alpha is? Thank you! $\endgroup$ – Ramalho Feb 15 '16 at 18:14
  • $\begingroup$ $\alpha$ is some real number in the given interval. $\endgroup$ – CKM Feb 16 '16 at 3:41
  • $\begingroup$ That I understood, but why do we need it? I understand that max is not differentiable by it self, but why multiply by some alpha?! I hope to understand. You are also missing the gradient with respect to $W_{yi}$. $\endgroup$ – Ramalho Feb 17 '16 at 10:23
  • $\begingroup$ Please look at the subgradient of binary class hinge loss here. $W_{y_i}$ is fixed; Not a variable since it is indexed by known class label. $r$ is the row/label that you choose; its variable. $\endgroup$ – CKM Feb 18 '16 at 5:07
  • $\begingroup$ P.S. If you think $W_{y_i}$ is variable, if you easily modify the above differentiation. $\endgroup$ – CKM Feb 18 '16 at 7:51

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.