3 ms·
I spent 3 years building a complex statistical model in Stan, only to double its performance with a trivial linear regression. The growth is realizing that som
by marmaduke 7y ago
I spent 3 years building a complex statistical model in Stan, only to double its performance with a trivial linear regression.
The growth is realizing that sometimes variance in the data means that really cool model/code you want to write isn't useful, and there's always a trade-off between elegance and utility (if the two aren't in fact opposites)
- aguzzi94 7y agoWow that's harsh! How did you feel when you realized it?
- marmaduke 7y agoRelieved actually, because it allowed to meet an overdue performance deadline.
- mojomark 7y agoHa, this is a great reflection. In a different domain I had a similar experience - working for about 2 years on an unmanned vehicle design I thought was revolutionary, only for a colleage to review the design and point out that reverting to a traditional propulsion engine increases performance by a factor of 2. I felt like a balloon deflating; alas life goes on. I suppose the lesson learned is avoid working in a vacuum of a project for long stretches and instead ask for a variety of consult from colleagues early and regularly.
- marmaduke 7y agoThat's a good remark as well: we didn't have effective iterations with feedback, just long arcs of speculative changes.
- cuchoi 7y agoWhat do you think was the cause of the underperformance of your Stan model?