Ex. 11.1

[szcf-weiya]

[szcf-weiya]

[ggg5269]

TKS but why choose $\alpha_{00}$ ?

[Huyang0625]

What's your loss function? and it doesn't make sense $g_m$ is dependent on k

[szcf-weiya]

What's your loss function? and it doesn't make sense $g_m$ is dependent on k

Thanks for your question.

Firstly, I am not sure why you care about the loss function? Secondly, I am sorry the notation here is not very clear, but

• For the regression, actually only 1 output unit, so no dependence on k.
• For the classification, which can turn out to be multi-response regression, so for each k, we still can establish the equivalence, i.e., for K classification problem, we can get K equivalent PPR models.

[Huyang0625]

What's your loss function? and it doesn't make sense $g_m$ is dependent on k

Thanks for your question.

Firstly, I am not sure why you care about the loss function? Secondly, I am sorry the notation here is not very clear, but

• For the regression, actually only 1 output unit, so no dependence on k.
• For the classification, which can turn out to be multi-response regression, so for each k, we still can establish the equivalence, i.e., for K classification problem, we can get K equivalent PPR models.

Thank you for your reply!

Your answer about the k-classification is helpful!

But I'm still confused how to deal with the learning process of the K-1 PPR, because for different k, $g_m$ share parameters like $alpha$ and which means we can't do the learning for K-1 models respectively. Can we still call it a PPR problem?

[szcf-weiya]

@Huyang0625 replied in http://disq.us/p/264iotd, and update the solution.