昨晚發過去的，今天一早就給出了反饋信息。可能是之前師兄做得差不多了，所以只有兩點針對證明的意見，下面是Wei_Fan的回信：

Some clarifications need to be made to the formal analyses:
 
1. It is not clearly to me, in Theorem 1, what exactly is b. This needs to be defined formally and clearly. I think that this is the bayesian optimal decision.
 
2. I do not understand why conf(x) = max p(y|x)?
 
     In reality, isn't the confidence of a prediction the estimated probability by a model M, and there is always a dependency on M? In other words, the estimated probability by a model is p(y|x,M) and there is an explicit dependency on M, and normally P(y|x) = P(y|x,M).
 
   I think that there is one more step need to be done. That is to assume M is better than random guessing, thus, P(y|x,M) is reasonably close to P(y|x),..