Machine learning is being used more and more nowadays, from machines to automatically process bank cheques to Google’s reCAPTCHA. So I recently spent some time learning about neural networks, and their application to machine learning and image recognition. I would like to recommend the book I found on this subject, Neural Networks and Deep Learning by science writer Michael Nielsen. The book is freely available online, in a web-only format.
The book covers the basic theory of how neural networks operate from the individual “neurons” to the different network topologies available including feedforward and recurrent neural networks. Only feedforward neural networks are covered in detail here, since these are the type most commonly used for image recognition. Later on, the process to train neural networks is covered in depth, including detailed explanations of the backpropagation algorithm and stochastic gradient descent. The final chapter, covers the techniques used to achieve deep learning in many layered neural networks and discusses some of the current limitations of neural networks and their future.
Throughout, the MNIST database of handwritten digits is used to train and test the neural nets developed in the accompanying code examples. The earlier examples are built with Python and NumPy, the deep-learning example uses the Theano library and can be run on a GPU. The MNIST database is a standard benchmark used to test image recognition accuracy with neural networks and other machine learning techniques such as support vector machines.
Overall, I would highly recommend this to anyone studying this subject at university or anyone with an interest in learning more about machine learning in general. Many of the techniques described in this book are applicable to other machine learning algorithms.