Modification of a neural network in Caffe to handle multiple label/category classification

We need to have a standard convolutional neural network[ either Alexnet or Googlenet] modified to handle multiple label/category classification. We intend to train it with the standard Imagenet type dataset with one labeled category object per image[for example one image with a dog, one with a tree, and one with a cat]. As those nets are currently designed if they exposed to a test image with objects from more than one category or class [for example both a cat and a dog] the output of the net is that the single most likely class is selected and the other classes/categories are inhibited and the probabilities normalized [cat 75%, dog 20%, tree 5%]- [see uploaded example of both a dog and a cat in an image but the probabilities favor the dog and the cat probability is suppressed although both are present]. What we need to be able to do is [after training the standard way as above] to expose the net to test cases with objects from more than one category. We would like to have each class that is present in the image listed with independent probabilities[ for example cat 90%, dog 90%, tree 03%]. Some approaches to this are described in the following Stackoverflow articles. [login to view URL] [login to view URL] One approach seems to be [from the second article]: "What we need: the problem with the existing approach is the "Softmax" layer that basically selects a single class. I suggest we replace it with a "Sigmoid" layer that maps each of the C outputs into an indicator whether this specific class is present in the image. For training, we should use "SigmoidCrossEntropyLoss" instead of the "SoftmaxWithloss" layer." I t appears that these changes can be made by simple modification of the last layer or so of the net. Unfortunately this is beyond our skill set and we need help implementing it. Deliverables: Alexnet or Googlenet network file modified to handle multiple label/category classification