Zeuter wrote:I missed over half of the lectures, studied for 3 days for an exam covering 2 months of machine learning, and aced it. There just isn't much to it anymore.
Academic training in this area appears to be worthless in practice. I work with a machine learning Ph.D. from Stanford and have had to coach him on the most basic things. For example we have a pretty standard SGD SVM + random kitchen sinks classifier in one product, and I asked him to apply Platt scaling to calibrate the posterior probabilities. Rookie shit. After a few hours I check in and he's stuck because the test model he trained returns 100% confidence in the same class every time. A few seconds looking at the code shows he hasn't scaled the features ... and it's not that he forgot. Oh no, that would be forgivable. It's that he, a machine learning Ph.D. from Stanford, doesn't already know that you need to scale features for an SVM, and has no intuition about why such a thing might be prudent for an SVM as opposed to, say, decision trees. Oy.
Heh, I've just gotten to SVMs, currently I'm working on training a small variety of models on the mnist handwritten digit dataset. No idea what random kitchen sinks are, scaling is a given though so that's hilarious.
I do agree that academic training doesn't map particularly well to real-world work, as you don't have the processing power, nor is it about making any particular accurate model. There's too much to cover in too little time, so I'm forced to do a broad analysis of how well different very basic algorithms (multinomial logit models, svms, neural nets) work in practice, as well as learning how everything theoretically works under the hood.
Re: games, although the tech can be interesting and applicable in other fields, it really is a terrible choice. It's all about money and power, and there is no money in indie games. If you cannot force the hand of your employer you will be exploited, and you will be unhappier for it. I'm considering migrating, as EU compensation is generally terrible, and the work is for the most part metaprogramming-less CRUD codemonkeying. You know, the kind of work that is outsourced to India. Frankly I think it's absurd that that work even exists, because if the foundations of CRUD shops were better, then there would be no work to outsource. There's plenty of companies and bizarre recruiters here trying to get me to work for them (ever read a recruitment email that read like it was dystopian scifi?), but that kind of work is just ridiculous.
However, the bay area startup scene seems mostly like a collection of fraudulent investment vehicles. Entrepreneurship in the EU is an option, but the quality (and volume) of investors in Europe is much, much lower than it is in the US. Any suggestions?