This is similar to the debate within the machine learning community around neural networks: many people really don't like that they are black boxes, and prefer other machine learning methods that can be reasoned about mathematically (for example have a rigorous probabilistic interpretation).

My personal feeling is that symbolic methods, and also to some extent the other genres of machine learning, in the end are not powerful enough to get to the "next level" of machine intelligence. I think the really interesting things arise as "emergent properties" from complex networks, these are inherently things which are impossible to reason about mathematically. Naturally, this is also a bit scary, and getting the networks right is more of an art than a science :-) The biological neural network (i.e. our brain) is the only existing implementation of an intelligent system, so clearly there is something to it :-)

BTW, neural networks are also kind of a an old thing resurfaced. It's a field of research that has had several "Dark Ages", and then someone has discovered how to do things better and there has been a strong resurgence. The most recent is the deep learning boom, which has caused neural networks to be popular again. (When I started research at the early 2000's it was just at the time of neural networks going out of fashion, since e.g. support vector machines were much superior. I stuck to them for a long time until finally giving up and doing my doctoral thesis on another topic... just before they came back again :-)

sazius at 2016-07-19T06:42:57Z

jrobertson, Claes Wallin (韋嘉誠), Christopher Allan Webber likes this.

@sazius Cool! Btw, I am not discounting neural networks... probably really good systems in the future will be a hybrid of systems anyway, and it's exciting that neural networks are now at the point that they can sometimes be like a better human intuition for certain well trained problems.

But I suspect while that branch of CS is getting a lot of attention right now, there's probably a lot of valuable, interesting things in this other branch that's not being as well paid attention to any more :)

Christopher Allan Webber at 2016-07-19T13:00:06Z

Claes Wallin (韋嘉誠) likes this.

I agree 100% that one should not focus solely on the most exiting branch of a science. This is why I'm opposed to the current philosophy of politicians here in Finland (and probably elsewhere too) that in order to save money we should focus research on only the most "promising" fields.

In fact, when neural networks were not considered as promising (e.g. early 2000's), there were still some people working on it, and some of them made the breakthroughs that enabled the current deep learning boom. You can easily find even more startling examples in other sciences... we need to have people exploring every branch, you never know where the next breakthrough will be :-)

sazius at 2016-07-19T17:36:54Z

Tobias Diekershoff, Tyng-Ruey Chuang, Charles Stanhope, Elena ``of Valhalla'' and 2 others likes this.

still would love to find better ways to find the most likely dates and cities for events in feeds users submit here

often the needed info is obvious to humans but rather hard to extract for a machine .. .. the trouble is that that machine doesn't have those many extra cultural hints of relatedness that humans are able to take into account in an instant without even noticing - theres a lifetime of learned connections in every human :-)

one thing that did quite well was to index lots of old listings data with a fulltext search application and scoring the results of searches. I was trying that with kinosearch (and its sccessor: Lucy) because you could though a lot of terms at it and get nicely weighted/ranked results pretty quickly and was easy to install and get started with and not too heavy to use on my old pc.

sure using searcxh tools that way was a pretty rough and hacky experiment, but for what it was it did better than i expected.

a yacy index could probably work too

played with lots of other methods but at the end of the day to do much good it still usually ended up needed more and more data so with my limited resources pretty much everything ended up turning into some kind of search experiment anyway because things would end up spending too much time waiting for data I guess a server farm or access some kind of shared grid thing would probably be wanted ... but maybe a few gpu's currently just sitting there idle could be used for something?

Michael at 2016-07-20T08:07:15Z