The projects covered in this book can be considered bite-sized projects. They can be completed within a day or two. A real project will often take months. They require a combination of machine learning expertise, engineering expertise, and DevOps expertise. It would not quite be feasible to write about such projects without spanning multiple chapters while keeping the same level of detail. In fact, as can be witnessed by the progression of this book, as projects get more complex, the level of detail drops. In fact, the last two chapters are pretty thin.
All said and done, we've achieved quite a bit in this book. However, there is quite a bit we have not covered. This is owing to my own personal lack of expertise in some other fields in machine learning. In the introductory chapter, I noted that there are multiple classification schemes for machine learning systems and that we'd be choosing the common view that there are only unsupervised and supervised types of learning. Clearly, there are other classification schemes. Allow me to share another, one that has five classifications of machine learning systems:
- Connectionist
- Evolutionary
- Bayesian
- Analogizer
- Symbolist
Here, I use the term machine learning. Others may use the term artificial intelligence to classify these systems. The difference is subtle. These five classes are technically schools of thought within artificial intelligence. And this sets a much larger stage for the topics at hand.
Except for two, we have, in this book, explored the different schools of thought in artificial intelligence. In the Connectionist school, we started with linear regression in Chapter 2, Linear Regression – House Price Prediction, and the various neural networks from Chapters 8, Basic Facial Detection, and Chapter 10, What's Next?. In the Bayesian school, we have Naive Bayes from Chapter 3, Classification – Spam Email Detection, as well as the DMMClust algorithm in Chapter 6, Neural Networks – MNIST Handwriting Recognition; we also have the various distance and clustering algorithms, which somewhat fall into the analogizer school of thought.
The two schools of thought on artificial intelligence that are not covered are the Evolutionary school and the Symbolist school. The former I only have theoretical experiences of. My understanding of the Evolutionary school of artificial intelligence is not great. I have much to learn from the likes of Martin Nowak. The latter, I am familiar with—I have been told that my introduction to Go betrays a lot of my experience with the Symbolist school of thought.
The main reason why I didn't write anything about the Symbolist school of thought is that as a subject matter it is too dense, and I am not a good enough writer to actually tackle the subject. It opens up hairy philosophical implications more immediately than the Connectionist school does. These implications are something I am not yet ready to deal with, though the reader might be.
Having said that, one of the most exhilarating times in my life was building DeepMind's AlphaGo algorithm in Go. You can find the code here: https://github.com/gorgonia/agogo. It's a behemoth of a project, and successfully pulled off by a small team of four. It was an immensely rewarding experience. The AlphaGo algorithm merges Connectionist deep neural networks with Symbolist tree search. Despite pulling off such a feat, I still do not think I am ready to write about the symbolic approach to artificial intelligence.
All of this brings up the question: what's next?