It seems that deep neural networks are making improvements largely because as we add nodes and connections, they are able to put together more and more abstract concepts. We know that, starting from pixels, they start to recognize high level objects like cat faces, chairs, and written words. Has a network ever been shown to have learned a more abstract concept that a physical object? What is the "highest level of abstraction" that we've observed?
You can train DNN to
learn compute any abstract concept just by making that abstract concept as the label (output) in the training dataset. For example there are projects which detects emotions from peoples photos.