Once the basic machine learning algorithms have been sufficiently covered, I’ll move on to more advanced algorithms and techniques.
Data science is a growing field with a lot ground to cover, so instead of trying to be comprehensive and risk analysis paralysis, here’s an initial starting list which I’ll expand over time:
- Advanced algorithms: Support vector machines, random forests, deep learning networks.
- Techniques: Bagging, boosting, ensemble, regularisation, cross validation.
Guiding Principles
It’s quite easy to get lost in the details and focus on model building, but the real value of using machine learning is to solve real-world problems by convincing decision makers with authority over resources to take action. In order to be able to convince them, I’ll be guided by some core principles.
- Keep it simple, stupid
Which means letting the problem guide the choice of solution, and starting with simple solutions and moving on to more complicated ones only when necessary. There’s no need to use a ten layer convolutional neural network when logistic regression would work just fine.
Don’t know what I mean? Maybe this GIF will help.
- Results, and how they were achieved, must be explainable
Which means minimising black box approaches, and when they really have to be used (deep neural networks come to mind), strive to explain how the results were achieved instead of saying “just trust the machine”. If you were on the receiving end of that assurance, would you?
- Human-plus-machine, not human-vs-machine
Which means augmenting machine learning with human intuition, abstraction and imagination. I believe that there will come a time, hopefully after I’m long gone, that true general artificial intelligence will emerge. And when the Singularity happens, I’m sure (hope?) our robot overlords will look kindly upon a long history of cooperation.
- Models degrade over time
Which is probably best exemplified by the rise and fall of Google Flu Trends. The world is fluid and models that are built at one point in time will inevitably degrade as the world changes. Therefore, models need to be continuously monitored to ensure that they remain relevant and accurate.
On the specific topic of deep learning, I find that “Deep Learning” by Ian Goodfellow et al has a good balance of breadth and depth.
Part I covers applied math and machine learning basics, Part II covers deep networks modern practices and Part III covers deep learning research i.e. the new stuff.
I’ll go through the 800 pages, and I’m sure that it will be time well spent.
There’s an old Chinese saying: “工欲善其事, 必先利其器”, which loosely translates to: “In order to a good job, you must first sharpen your tools”. Wise words.
