The Galaxy Zoo Kaggle Competition
Back in February, a few colleagues and I decided to go for our first Kaggle competition! We entered the Galaxy Zoo challenge to classify galaxy morphologies. For some of us, this was the first time doing machine learning. And many of us hadn’t done any image processing or feature extraction before.
One of my colleagues, Kwyn Meagher, documented our process through two blog posts - Machine Learning Work Flow for Hackers and So You Want to Be a Kaggle Wizard. I took the reins with implementing a random forest algorithm, choosing parameters, and testing features. We ended up in 120th place out of 329 teams. Not bad for our first Kaggle competition.
The Winning Solution
I’m very happy that the winner, Sander Dieleman, has decided to post his code as well as a detailed explanation of how he approached the problem. This kind of generosity within the machine learning and Kaggle community is priceless.
It’s interesting to see that the top two people on the leaderboard both used convolutional neural networks in their solutions. Having just finished implementing my first neural network learning algorithm today, this is more than inspiring. I’m really looking forward to diving into his explanation in some detail.