In this final chapter, we moved some distance from our previous discussions of algorithms, configuration and diagnosis to consider tools that improve our experience when implementing deep learning algorithms.

We discovered the advantages to using Lasagne, an interface to Theano designed to accelerate and simplify early prototyping of our models. Meanwhile, we examined TensorFlow, the library developed by Google to aid Deep Learning model adjustment and optimization. TensorFlow offers us a remarkable amount of visibility of model performance, at minimal effort, and makes the task of diagnosing and debugging a complex, deep model structure much less challenging.

Both tools have their own place in our processes, with each being appropriate for a particular set of problems.

Over the course of this book as a whole, we have walked through and reviewed a broad set of advanced machine learning techniques. We went from a position where we understood some fundamental algorithms and concepts, to having confident use of a very current, powerful and sought-after toolset.

Beyond the techniques, though, this book attempts to teach one further concept, one that's much harder to teach and to learn, but which underpins the best performance in machine learning.

The field of machine learning is moving very fast. This pace is visible in new and improved scores that are posted almost every week in academic journals or industry white papers. It's visible in how training examples like MNIST have moved quickly from being seen as meaningful challenges to being toy problems, the deep learning version of the Iris dataset. Meanwhile, the field moves on to the next big challenge; CIFAR-10, CIFAR-100.

At the same time, the field moves cyclically. Concepts introduced by academics like Yann LeCun in the 80's are in resurgence as computing architectures and resource growth make their use more viable over real data at scale. To use many of the most current techniques at their best limits, it's necessary to understand concepts that were defined decades ago, themselves defined on the back of other concepts defined still longer ago.

This book tries to balance these concerns. Understanding the cutting edge and the techniques that exist there is critical; understanding the concepts that'll define the new techniques or adjustments made in two or three years' time is equally important.

Most important of all, however, is that this book gives you an appreciation of how malleable these architectures and approaches can be. A concept consistently seen at the top end of data science practice is that the best solution to a specific problem is a problem-specific solution.

This is why top Kaggle contest winners perform extensive feature preparation and tweak their architectures. It's why TensorFlow was written to allow clear vision of granular properties of ones' architectures. Having the knowledge and the skills to tweak implementations or combine algorithms fluently is what it takes to have true mastery of machine learning techniques.

Through the many techniques and examples reviewed within this book, it is my hope that the ways of thinking about data problems and a confidence in manipulating and configuring these algorithms has been passed on to you as a practicing data scientist. The many recommended Further reading examples in this book are largely intended to further extend that knowledge and help you develop the skills taught in this book.

Beyond that, I wish you all the best of luck in your model building and configuration. I hope that you learn for yourself just how enjoyable and rewarding this field can be!