:quality(75))
- The post describes using H2O to build a Kaggle competition submission for the MNIST dataset while following the leaderboard for the ongoing contest.
- The post reports H2O built a model of 500 trees on a server with 20 gigabytes of memory in about the time it took to grab takeout and coffee.
- The post states the resulting Kaggle submission achieved less than 5 percent prediction error on the digit recognition task.
- The post frames H2O's purpose as removing the need to spend days finding clever, narrowly generalizable tricks for a single dataset.
Since we've been fooling around with the MNIST data set quite a bit lately (Spence is using it in benchmarking), I've been following the leaderboard and methods for the ongoing Kaggle competition around the same data. It's really amazing to see what people come up with. But of course, the purpose of H2O is entirely that one need not devote days on end to finding clever tricks with limited generalizability to other predictive problems. So while Spence benchmarks speed, I made short work of a submission that let H2O do all of the heavy lifting, to get a sense of how we're doing on predictive power for real. In the time it took me to grab take-out and coffee H2O built a a model of 500 trees (on our server with 20g memory. It would take a bit longer on my computer with an allocation of less memory), generating a submission with less than 5% prediction error – all in all, not a bad lunch break.