Questions on target distribution and others of lecture four

[LazyPiggy]

I don't understand the target function mentioned in slide 16, why the probability of y given x is used to represent the target distribution instead of the probability of y alone? so is the deterministic target.Another question is that what is the smoothness of square error curve affects the error measure? The last one is that is there any empircal principle for choosing error measure alternatives, namely, in which cases plausible measures are more superior to friendly ones? All comments are appreciated.

[yaser]

Let me quickly answer the first question. The target, whether function or distribution, is the dependency of on . That is what we are trying to learn. If that dependency is deterministic, i.e., uniquely determines , we have a target function. If that dependency is not deterministic and we have different 's that may occur when occurs, we have a target distribution . The fact that it depends on is inherent, as a different would result in a different distribution of 's.

[LazyPiggy]

Originally Posted by yaser

Let me quickly answer the first question. The target, whether function or distribution, is the dependency of on . That is what we are trying to learn. If that dependency is deterministic, i.e., uniquely determines , we have a target function. If that dependency is not deterministic and we have different 's that may occur when occurs, we have a target distribution . The fact that it depends on is inherent, as a different would result in a different distribution of 's.

Thank you for your intelligible answer. And what about the second and the third one?