Ask HN: Any real uses of Neural Networks? | Hacker News Reader

Ask HN: Any real uses of Neural Networks?

5 pointsTeMPOraL4 comments

I'm struggling to find any use case for neural networks, where they wouldn't be existing solutions that are better and more manageable. Any ideas where and in what context neural networks are actually useful?

View on HN

They're extensively used in particle physics as a method of performing multi-variate selections of data. It's a common case that we have a large data set which contains the signal we're looking for. Each entry in the data set will have a large number of variables, some of which will provide some discriminating power between our signal and background. By training a neural network to use those variables to tell the difference between signal and background we can extract the maximum efficiency from our data.

Of course we could use any other multi-variate method (boosted-decision trees are popular) but neural networks are becoming more and more common.

They're a reasonable supervised learning technique. I don't prefer them myself, but they have some advantages for online learning and low memory requirements, among other things, and do decently in supervised-learning-roundup benchmarks (e.g. http://www.cs.pitt.edu/~litman/courses/cs2710/papers/empiric...). Of the older supervised-learning techniques (before SVMs/boosting/etc.), they tend to do better than alternatives such as decision trees (though decision trees are more interpretable).

While this is more anecdotal, I know quite a few financial firms use neural networks for their stock trading algorithms. It's less about whether or not they're the right tool, and a matter of using every tool there is to see what happens to work best. Here's a very brief example: http://www.i2r.org/nnstocks.pdf

Also, often times these methods are combined to get even better results. See "ensemble learning: http://en.wikipedia.org/wiki/Ensemble_learning

I used growing neural gas ( http://web.cs.swarthmore.edu/~meeden/DevelopmentalRobotics/f... ) for my MS research dealing with keystroke dynamics (specifically, modeling Hidden Markov Models on interkey-delays to identify a user's language). When you have a large set of data and you need to identify exemplars for that data, it does a pretty good job. As _delirium said, though, it requires a degree of supervision and tweaking to get growth parameters, etc.

Edit: I use 'supervision' in a 'it might take a few tries to get it working ideally' sense, not a 'you need a priori knowledge of the data' sense.

Reply on news.ycombinator.com