Neural networks are leading the current machine learning trend. Whether it's Tensorflow, Keras, CNTK, PyTorch, Caffee, or any other package, they are currently achieving results that few other algorithms have achieved, especially in domains such as image processing. With the advent of fast computers and big data, the neural network algorithms designed in the 1970s are now usable. The big issue, even a decade ago, was that you needed lots of training data that was just not available, and, at the same time, even when you had enough data, the time required to train the model was just too much. This problem is now more or less solved.

The main improvement over the years has been the neural network architecture. The backpropagation algorithm used to update the neural networks is more or less the same as before, but the structure has seen numerous improvements, such as convolutional layers instead of dense layers, or, Long Short Term Memory (LSTM) for regular recurrent layers.

Here is the plan that we will follow: first a deep dive inside TensorFlow and its API, then we will apply it on convolutional neural networks for image processing, and finally we will tackle recurrent neural networks (specifically the flavor known as LSTM) for image processing and text processing.

Talks about machine learning speed are mainly about neural network speed. Why? Because neural networks are basically matrix multiplications and parallel math functions—blocks that GPUs are very good at.