My previous blog post about deep learning for chess blew up and made it to Hacker News and a couple of other places. One pretty amazing thing was that the Github repo got 150 stars overnight. There was also lots of comments on the Hacker News post that I thought were really interesting. (See this skeptical comment for instance).
Read more…I’ve been meaning to learn Theano for a while and I’ve also wanted to build a chess AI at some point. So why not combine the two? That’s what I thought, and I ended up spending way too much time on it. I actually built most of this back in September but not until Thanksgiving did I have the time to write a blog post about it.
Read more…Say you build a machine learning model, like a movie recommender system. You need to optimize for something. You have 1-5 stars as ratings so let’s optimize for mean squared error. Great.
Then let’s say you build a new model. It has even lower mean squared error. You deploy it. This model turns out to give a lower mean squared error. You roll it out to users and the metrics are tanking. Crap! Ok so maybe mean squared error isn’t the right thing to optimize for.
Read more…I keep forgetting to buy a costume for Halloween every year, so this year I prepared and got myself a Luigi costume a month in advance. Only to realize I was going to be out of town the whole weekend. If anyone wants a Luigi costume, let me know!
I spent a couple of hours this weekend going through some pull requests and issues to Annoy, which is an open source C++/Python library for Approximate Nearest Neighbor search.
I set up Travis-CI integration and spent some time on one of the issues that multiple people had reported. At the end of the day, it turns out the issue was actually caused by a bug in GCC 4.8. Some crazy compiler optimization introduced between 4.6 and 4.8 caused this loop to be removed:
Read more…I’m at RecSys 2014, meeting a lot of people and hanging out at talks. Some of the discussions here was about the filter bubble which prompted me to formalize my own thoughts.
I firmly believe that it’s the role of a system to respect the user’s intent. Any sensible system will optimize for user’s long-term happiness by providing info back to the user that s/he finds useful. This holds true as long as a system isn’t (a) stupid and recommends the wrong content (b) trying to push its own agenda, that may or may not be hidden.
Read more…Note: This is a silly application. Don’t take anything seriously.
Benford’s law describes a phenomenon where numbers in any data series will exhibit patterns in their first digit. For instance, if you took a list of the 1,000 longest rivers of Mongolia, or the average daily calorie consumption of mammals, or the wealth distribution of German soccer players, you will on average see that these numbers start with “1” about 30% of the time. I won’t attempt at proving this, but essentially it’s a result of scale invariance. It doesn’t apply to all numerical series, like IQ or shoe size, but this pattern turns out to pop up in a lot of places.
Read more…Inspired by Sander Dieleman’s internship at Spotify, I’ve been playing around with deep learning using Theano. Theano is this Python package that lets you define symbolic expressions (cool), does automatic differentiation (really cool), and compiles it down into bytecode to run on a CPU/GPU (super cool). It’s built by Yoshua Bengio’s deep learning team up in Montreal.
Read more…Many years ago, I used to think that A/B tests were foolproof and all you need to do is compare the metrics for the two groups. The group with the highest conversion rate wins, right?
Read more…I’ve been spending quite some time lately playing around with RNN’s for collaborative filtering. RNN’s are models that predict a sequence of something. The beauty is that this something can be anything really – as long as you can design an output gate with a proper loss function, you can model essentially anything.
Read more…