The quarterly review — Q2
Winter 2017. Second quarter of grad school. The quarter was as cold and brutal as the weather. Coming from a coastal Indian city which never sees single digit temperatures, this quarter was completely in single degree temperatures, and despicable late evening classes.
A TL;DR; summary of this quarter should be coffee, project, more coffee, paper; sleep; homework, right from day 1. My first two weeks were spent auditing so many classes, that I had classes all week. I wasn’t sure what courses I wanted to take, even at the end of second week.
The first course I took was CSE 253 - Neural Networks. Little did I know how hectic this course was going to be. The programming assignment 1 was released the day before class started, and everything in this course, including the course contents moved way too quickly. I invested most of my week to this course, with the onslaught of programming assignments, midterms and final project, and the fact that the course content was challenging (in the span of 10 weeks). Nevertheless, in the span of 11 weeks,
- I wrote a MNIST digit classifier using logistic regression and neural networks
- trained a neural network to recognize the urban tribe,
- generate music using RNN
- detect duplicate question pairs on Quora. Woot!
I also took CSE 258 - Web Mining and Recommender Systems - A fascinating subject, and an almost sweet, well structured course. It started very slowly, with the basics of supervised and unsupervised learning. I put in the least amount of effort for the homeworks of this course. And suddenly, came the first Kaggle assignment, and it was a challenge. The research project was relatively okay; we worked on recommending new restaurants to users in Las Vegas using a cool technique called as Bayesian Personalized Ranking (BPR). I will be pursuing my research in this domain of recommender systems, and I am excited for the next quarter to start.
The third course I took was CSE 291 - Convex Optimization. I have never had to think so deeply in math. This course made me use up all my energy and made me rack my brains for all the linear algebra I knew. It was mentally exhausting, but a fascinating subject. I am just glad that I took the CSE offering of the course, instead of the ECE or the Math one, as I could rely on geometrical intuition for proofs, rather than hard mathematical rigor. It was not at all obvious as Stephen Boyd, the god of convex optimization puts it. In my last homework, I attached a page of a 1998 research paper as my answer for one question. Never have I felt so incredibly stupid, and yet fallen in love with the subject. I got to have another perspective of many things in machine learning, which has a foundation in convex optimization as well. Another aspect I liked in this course was the class size was 13. I got to interact with the professor a lot more than I could in other classes. I also wrote a review paper on algorithms for bandit convex optimization. I probably understood less than 40% of the course properly, but it was fun - especially the late night group brainstorming. In fact, our Whatsapp group name was “Sleep when it’s obvious”.
I am just glad that this quarter is over. It’s a relief not to jump from one deadline to other. Having taken 6 courses over 2 quarters, I don’t think I will ever take three courses in a quarter. It’s mentally debilitating. Phew. The quarter is done, and the spring break is on!