0
$\begingroup$

the following equations are given:

$\sum_{j=1}^c\hat{P}_j = 1$

$\sigma_i(\mathbf{z}; \theta) = \frac{exp(\mathbf{\theta}_i^T\mathbf{z})}{\sum_{j=1}^cexp(\mathbf{\theta}_j^T\mathbf{z})}$

$L = \sum_{j=1}^c \hat{P}_j \, log(\sigma_j(\mathbf{z};\theta))$

How can I prove the following?

$\nabla_{\theta_i}L = (\hat{P}_i - \sigma_i(\mathbf{z};\theta))\,\mathbf{z}$

This is from the following paper: http://icml.cc/2012/papers/389.pdf Equation number 19.

Thanks in advance.

$\endgroup$
1
$\begingroup$

$\textbf{hint}$

$$ \log(\sigma_j(\textbf{z};\theta)) = \theta_{i}^{T}\textbf{z} -\log\left(C(\textbf{z},\theta)\right) $$ where $$ C(\textbf{z},\theta) = \sum_{j=1}^{c}\mathrm{e}^{\theta_{i}^{T}\textbf{z}} $$

and $$ \bigtriangledown_{\theta_j}\log\left(C(\textbf{z},\theta)\right) = \dfrac{\bigtriangledown_{\theta_j} C(\textbf{z},\theta)}{C(\textbf{z},\theta)} $$

can you take it from 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.