Update nnbp.m

[ericstrobl]

added rectified linear and softplus activation functions

[rasmusbergpalm]

Please merge this into your other pull request, and see the comments there. Also I think happynears suggestion should work, so please try that.

[AK747]

Hello.I will ask a problem about convolutional neural network. There are forty-five gray pictures (256*256),I use them as traing sample, which is divided into three categoriesapple（apple, camel , car ）, each category 15 pictures: apple is set label 1, camel is label 2, car is label 3.Then I also put them into 100,010,001. I choose extra fifteen pictures as traing sample.My mat documents types are the same with handwritten database. I simply modify the main function data, while I haven't changed other functions. Regardless of the iteration number, the results are the same. Divided into three categories, the error rate is 2/3. Accordingly, I also test ohers, if divided into four categories, the error rate is 3/4;Divided into five categories, the error rate is 4/5...... I found the function of cnntest (~, h] = max (net.o) was wrong.I got the results are the same value, such as, all is 1 or 2 or 3 ...... Why this was happen? Why not get the right h?