The Softmax classifier does not leverage the full array of class-specific information encoded in a trained neural net. This paper describes a method to better classify difficult samples by using information ignored by Softmax.

A version is available on arXiv at money on the table.

You can get the PDF directly.