One aspect all recent machine learning frameworks have in common – TensorFlow, MxNet, Caffe, Theano, Torch and others – is that they use the concept of a computational graph as a powerful abstraction. A graph is simply the best way to describe the models you create in a machine learning system.