Machine Learning in a Year

My biggest regret is that my day job as developer does not leave me with much time for more Octave programming. Of course I do excercises, but that's not enough to be good ML developer.

However, I think that just having a lot of sample code would be helpfull if I ever make ML my full time job. Especially considering that the excercises themselves are pretty interesting (handwriting recognition, image compression, etc).

I'm taking the edX series of courses on Machine Learning with Apache Spark. It is a pretty good class and covers linear algebra and the basic ETL workflow, plus selection of models, model parameters tweaking parameters etc., study cases involving user ad clicks prediction and PCA analysis of neural synapse data of jellyfishes.

However, my biggest fear is that re-learning different matrix manipulations, walk-through of logistics regression, PCA and SVM will be similar to learning Spanish in high school without immersion, learning guitar scales without improvising with it, learning mappings of different Madden key-combo's; I remember learning all of these mathematical syntactical operations in high school, and have an eerie feeling of once grasping these concepts concretely once and yet only knowing it vaguely. It is nice to re-learn these things, like refreshing myself with a rolodex of Spanish verb conjugation, only to go back to the CRUD work I deal with on a daily basis and fade back to oblivion.

Not sure about the Coursera class, the structure of the edX courses are presented in Python Juypter notebooks where you fill in the code snippets. It is fun and addictive to solve each exercise as a mini-puzzle; but not sure how much it'll stick vs. if one had to take an concrete problem and take its pieces and puzzles from start to finish, without the hands-holding.

I think perhaps Cousera courses have "capstone" projects - curious if anyone have had experience doing one? In lieu of one on the edX course, I think I plan to grok some papers with some large genomic dataset and/or financial time-series and try to replicate their result.

Really, it's a great model for teaching technical subjects in general.

Above all have a good understanding of over fitting and best practices for test and validation sets. Then you can use algorithms as black boxes and at least know you aren't doing any harm.

EDIT: Note that it's still EXCEPTIONALLY difficult to provide true, lasting value to a given organization with this stuff and takes years of experience (note I didn't mention deep knowledge and experience with the latest techniques aboard the hype train).

Aside from that it's still trial and error. As a sibling pointed out the best you can do is get a good grounding in stats and sampling theory so that you ensure that you're not overfitting your data.

There is no rule-book for what kind of classifier works well on your problem, aside from reading the literature and seeing how well other people got on. There are exceptions, for instance if you're doing image classification then nowadays it's convnet or bust. But still you're going on instinct rather than "If you have problem <X> you should solve it with machine learning algorithm <Y>".

As with all these sorts of things, knowing the theory is useful if you want to start implementing your own classifiers from scratch, but nowadays it's approaching wheel-inventing territory. Machine learning libraries are mature, powerful and increasingly easy to use (like sklearn). Some of the libraries like Tensorflow are specifically geared so that you don't need to know about the implementation, you just read a description of the CNN and you can express it in a few lines of Python.

Otherwise using them is still pot luck:

- How many training samples will be enough for your problem to avoid overfitting/poor results? You won't know this until you try.

- What features are the best for your problem? There are algorithms which will try to learn for you, but in most cases that's a niche application and will require a lot of twiddling.

- You still (usually) have do a grid-search to find the optimal parameters for a particular method.

- Do you care which method you use, say an SVM vs Random Forest, if you get 90% accuracy consistently?

- Be mindful that for many problems in machine learning, 80% accuracy is considered exceptionally good. You shouldn't expect perfect performance so sooner or later you have to say good is good enough.

Only you can be the judge of that.

A better question is "What is the opportunity cost?". Is there something else that one could do with that same time that would be more valuable/useful?

I suspect that most persons have a number of things that would be more valuable to them than learning ML to some vague degree and undefined purpose. One might take an economics course, exercise more and pay more attention to one's diet to get into better shape, get more sleep, sell the boat, read Nicholas Taleb's second and third books, etc.

"Data Scientists Automated and Unemployed by 2025?" http://www.kdnuggets.com/2015/05/data-scientists-automated-2...

Nothing to get excited about, especially since you aren't yet invested. However, should the odds indeed be high that ML will become automated by 2025 then another field choice might be wiser.

How/where did you find a teacher like this? Finding a machine learning subject expert is already pretty difficult; knowing how to find one who is also a good teacher would be a useful skill in its own right.

Since your story is a success, no matter how you personally feel about it, can you tell us how many hours per week in that year did you dedicate your time to ML? A rough estimate will be good.

What are your plans for future? Find a ML-only job?

Would turning the results into a probability be useful? (e.g. "how likely are we to get a customer based on this description?") Your staff could generate test data for that.

Thanks, great article!

Very low math requirement, exercises in Python.

Depending on your learning style, Data Science from Scratch[2] might be another good option.

BTW, neither of these uses Octave like Andrew NG's course does. The first one uses R and the second uses Python.

[1]: http://www-bcf.usc.edu/~gareth/ISL/ [2]: http://joelgrus.com/2015/04/26/data-science-from-scratch-fir...

I think the problem with Coursera is that they try to accommodate as many students as possible, which means the classes aren't very deep. I don't think this should be the role of a ML class to teach about partial derivatives or matrix factorization for instance. It's great that the class is accessible to many people, but sometimes you feel that you just scratch the surface, and that you're not at the level of university students that went through more challenging classes.

As a comparison, I did the labs of 6.828 and 6.824 (OS and distributed systems) on MIT opencourseware and it was much harder and much more rewarding than any coursera class. And I did only the labs, there is much more material to cover to do it seriously (quizzes, lot of articles to read...). Kudos to MIT students! (I wonder how many such classes they take during one semester). Too bad opencourseware classes aren't always properly set up for outside students. There may be some material missing, no forum to ask questions, video quality can be very bad and so on...