We already saw some examples of using TensorFlow, and it's now time to understand more about how it works.

First things first, the name comes of the fact that TensorFlow uses tensors (matrices with more than two dimensions) for all computations. All functions work on these objects, returning either tensors or operations that behave like tensors, with new names defined for all of them. The second part of the name comes from the graph that underlies the data flowing between tensors.

Neural networks were inspired by how the brain works, but it doesn't work as the model use for neural networks. Yes, each neuron is connected to lots of other neurons, but the output is not a product of the input times a transition matrix plus a bias fed inside an activation function. Also, neural networks have layers (deep learning refers to neural networks with more than one so-called hidden layer, meaning neither the input nor the output) with a strict progression in their architecture. A brain has connections all over the place and continuous evolution, whereas a neural network always has a stable output for a given input and a given moment in time (until we get a new tick, as we will see in recurrent networks).

Let's now dive into the TensorFlow graph API.