This shows you the differences between two versions of the page.
project:hillclimb_dp:draft [2014/05/02 13:41] taolei [Justification of the Algorithm] |
project:hillclimb_dp:draft [2014/05/02 14:57] (current) taolei [Justification of the Algorithm] |
||
---|---|---|---|
Line 109: | Line 109: | ||
Unfortunately, it is not clear if there is an efficient way to compute this loss function by finding $y$ and $y'$. Instead, we could use the standard structural learning loss function, which corresponds to condition (1): | Unfortunately, it is not clear if there is an efficient way to compute this loss function by finding $y$ and $y'$. Instead, we could use the standard structural learning loss function, which corresponds to condition (1): | ||
$$ | $$ | ||
- | \mathcal{L}(\hat{x_i},\hat{y_i}) = S(\hat{x_i}, \tilde{y}) + \text{err} (\tilde{y}) - S(\hat{x_i}, \hat{y_i}) \quad\quad (1o) | + | \mathcal{L}(\hat{x_i},\hat{y_i}) = \max_{\tilde{y}}\left\lbrace S(\hat{x_i}, \tilde{y}) + \text{err}(\tilde{y})\right\rbrace - S(\hat{x_i}, \hat{y_i}) \quad\quad (1o) |
$$ | $$ | ||
where $\tilde{y}$ is the output of the decoding algorithm. | where $\tilde{y}$ is the output of the decoding algorithm. |