Softmax Regression
From Ufldl
(→Mathematical background) |
|||
Line 49: | Line 49: | ||
\begin{align} | \begin{align} | ||
\frac{\partial \ell(\theta)}{\partial \theta_k} &= \frac{\partial}{\partial \theta_k} \ln \theta^T_{y^{(i)}} x^{(i)} - \ln \sum_{j=1}^{n}{e^{ \theta_j^T x^{(i)} }} \\ | \frac{\partial \ell(\theta)}{\partial \theta_k} &= \frac{\partial}{\partial \theta_k} \ln \theta^T_{y^{(i)}} x^{(i)} - \ln \sum_{j=1}^{n}{e^{ \theta_j^T x^{(i)} }} \\ | ||
- | &= I_{ \{ y^{(i)} = k\} } x^{(i)} - \frac{1}{ \sum_{j=1}^{n}{e^{ \theta_j^T x^{(i)} }} } e^{ \theta_k^T x^{(i)} } \\ | + | &= I_{ \{ y^{(i)} = k\} } x^{(i)} - \frac{1}{ \sum_{j=1}^{n}{e^{ \theta_j^T x^{(i)} }} } e^{ \theta_k^T x^{(i)} } \qquad \text{(where } I_{ \{ y^{(i)} = k\} } \text{is 1 when } y^{(i)} = k \text{ and 0 otherwise) } \\ |
- | &= I_{ \{ y^{(i)} = k\} } x^{(i)} - P(y^{(i)} = k | x^{(i)}) | + | &= I_{ \{ y^{(i)} = k\} } x^{(i)} - P(y^{(i)} = k | x^{(i)}) |
\end{align} | \end{align} | ||
</math> | </math> | ||
With this, we can now find a set of parameters that maximises <math>\ell(\theta)</math>, for instance by using gradient ascent. | With this, we can now find a set of parameters that maximises <math>\ell(\theta)</math>, for instance by using gradient ascent. |