Saturday, December 8, 2018

Everything You Wanted to Know About Machine Learning....

...but didn't know whom to ask! 

Prof Sanjeev Arora, Princeton University, speaking at the Institute of Advanced Study

In a simple one-hour lecture (although one which does assume a certain amount of mathematical sophistication on the part of the audience, which, given that the audience consists mostly of mathematicians at the Institute of Advanced Study, is amply justified!), and one which already has had several thousands of views on youtube, Professor Arora describes the three main categories of Machine Learning Algorithms (Supervised, Unsupervised and Reinforcement Learning), and provides recent examples of 'successes' in each. Reinforcement Learning, which is taken up last, is nevertheless covered in sufficient detail, not leaving the viewer disappointed, and somewhat to one's  pleasant surprise as the lecture unfolds!

One of the topics taken up in some detail is the prediction of ratings for books based on the text of existing reviews, an application under the broad rubric of Natural Language Processing - and a subject of Prof. Arora's own recent work. The topic well illustrates many of the practical issues and methods involved in applications of machine learning, and also provides the intuition for developing theoretical ideas regarding how machine learning works. My own perspective is that of someone who knows much of the subject matter of the lecture for nearly three decades, since the mid-1980s, but I nevertheless found it a useful hour well-spent (on which, one might also add, that the lecture finishes almost exactly in an hour, showing also excellent time management on Prof Arora's part, as the lecture dynamically evolves!). What I greatly appreciated also was that Prof Arora reiterates several times during the talk the main reasons for the heightened interest in machine learning today, and this gives further insight into why someone - even if they might have been, for whatever reason (but especially out of an informed skepticism regarding such methods, as Prof Arora hints was the case with him) previously indifferent to the field - might now become interested, or even very interested in it. Another aspect of the lecture worth mentioning, especially these days when it is such a rarity, is that Prof Arora uses blackboard-and-chalk throughout - this added to my own enjoyment of the lecture, and doubtless also did the same for its live audience - a certain spontaneity is inevitably lost when lecturers use slides prepared earlier.

Two thumbs up, then!