We need advice on object recognition and labeling. Specifically we have two objects in an image (sometimes more) and those objects are attached. We want the neural net to recognize both objects and that they are attached even if it can't see the attachment location. We are looking for advice on machine learning and methods of training. We want advice on how to improve results using neural nets or maybe other methods. Our current attempts center around Hinton's deep neural networks but we are not getting good results right now.