Ask HN: Is AI a growing field? | Hacker News Reader

Ask HN: Is AI a growing field?

28 pointsdmix43 comments

I have become interested in artificial intelligence lately and I'm trying to gauge the field, but I haven't found many publications about the current state of it.

Are more people becoming interested in AI? Is there greater demand for it in businesses then say 10 years ago?

View on HN

If by AI you mean machine learning, yes. If you mean "human-like-interactive-thing," yes but not in a way that means we're substantially closer than we were 15 years ago.

AI in the ML sense is quickly becoming critical to business and is opening up lots of possibilities that we could only dream of 5 years ago. At its heart, Google is basically an ML company for one big example. Facebook is moving in that direction as well. There are lots of startups solving old problems with AI, too, such as Knewton (tutoring), and for some self-promotion, my company (Quantios).

Google: "Did you mean: Quantos"

We'll overshadow that eventually :) That's an unfortunate side-effect of starting out now; Quantios was in part chosen by virtue of domain-name availability, so "Did you mean..." is a predictable side-effect of that while we get off the ground.

Computer vision seems to be in high demand, too. The army wants its robots to know what they're shooting at.

Yes, they do. Reminds me of the AI algorithm designed to tell US tanks from Soviet ones. They accidentally trained it on one type of film for US tanks, and another type for Soviet tanks, each with different color tones, and ended up with a system that could spot different color tones...

http://books.google.com/books?id=JNhRSAMFn6YC&pg=PA271&#...

That's one of the problems with neural nets...there's usually no easy way for a human to understand what the actual decision process is (in sharp contrast to, say, decision trees).

Yes there is. There are a few methods of converting a trained nn to a decision tree. It's actually a better way of training decision trees for some problems.

Can you tell me what exactly it is that Facebook does with Machine learning?

I'd say no. My dad worked in AI 20-30 years ago(worked on some AI project for Russian military, something to do with nukes), and he looks for jobs in the field, now and then. And from what he say about, it seems like the AI field is pretty much frozen right now.

There are almost no jobs available, and the few available are usually the be assistant to some professor who is trying to do something that was already done 20 years ago.

Had the same impression. The most interesting advances I saw when visiting Alife XI last year was in cognitive robotics. Using MRI scan information from thinking persons to unravel the global structure of the brain, and then implementing that in computers.

The global workspace theory definitly seemed interesting.

I'm currently going around visiting potential Ph.D. programs as an admit for next year, and at all the top schools (CMU, Stanford, Berkeley) the amount of people studying AI and machine learning is staggering. The field is definitely academically very active.

There are a LOT of publications on AI if you look at all the conference papers and journal papers floating around on the net. The field of AI itself is divided up into lots of sub fields such as logic theory, machine learning, language processing, search/optimization.

Examples of "AI research at work" are the face detection in google picasa/iphoto ( machine learning + vision ). You email spam filter ( text classification ).

As AI research matures and as computers get faster, its safe to say that there is definitely a greater demand for AI but its hard to come up with "obvious" applications of AI algorithms like the ones I listed above.

as peter norvig said at a recent conference, it depends on how you define AI.

to gauge the field, why not start with his textbook (http://aima.cs.berkeley.edu/). right next to this link, i just saw this one here for more resources on the web (http://aima.cs.berkeley.edu/ai.html).

btw, i wouldn't really think too much about what other people are interested in or not if you are at your stage. if you are drawn to the field, do it. tech is too unpredictable to make this kind of career calculus.

Depends on what you mean by AI. It's not a single area of study.

Exactly. Additionally, the AI label is changing its meaning over the time.

My AI professors used to tell semi-jokingly that whenever we get computers to solve some task we thought required intelligence (thus we produced AI), the "intelligence" part of the AI gets redefined to exclude this particular solved topic (e.g. chess). We humans need to feel we are special.

Keeping this in mind, AI in the narrow sense is stagnating, AI in the broad sense is growing.

You could see "AI field" as a kind of incubator: cool stuff that works gets splintered into own subfields (computer vision, robotics, machine learning, neural networks, fuzzy systems, evolutionary algorithms etc).

----

Also, Wikipedia seems to have quite nice overview of the history of AI:

http://en.wikipedia.org/wiki/History_of_artificial_intellige...

Somewhere I heard a saying that goes something like "AI is machine intelligence we don't quite understand. Once you've figured it out, it is no longer AI, it is an algorithm."

I don't find that especially convincing - there is no machine intelligence we don't understand, since we have created all of it.

I would say that AI includes a few classes of algorithms related to solving certain types of problems (detection, planning, concept abstraction), or perhaps related to certain techniques for solving problems (heuristic search, ML, vision, etc.).

If your only talking about machine intelligence that we've already created, then of course we already understand that. I was talking about the concept of machine intelligence and problems that we would like to solve and are trying to solve with machine intelligence, but haven't solved yet. Once we solve a machine intelligence problem, we tend to call it an algorithm or a technique rather than machine intelligence.

Try searching AGI or "strong AI", which is a new movement trying to achieve general AI rather than application specific. I have got a good article on that: http://www.cybertechnews.org/?p=689

Don't go into THAT field unless you have guts of STEEL.

Either that or believe in the singularity.

The singularity is probably inevitable. I doubt it we'll see it in our lifetimes, though.

Why? If you buy into exponentially increasing progress, then most of the graphs seem to indicate that it will happen within a lifespan from now, especially if you consider the medical advances that will come in the intervening time.

All real-world exponential curves I can think of eventually reverse precipitously, or at least plateau.

This is mentioned in Kurzweil's Singularity book. With Moores Law, he expects transistors to peak at 2020 and molecular computing to take over after that and continue the growth.

He believes innovation and paradigm shifts supplement the plateauing of S-curves.

I'm saying even if you do buy any of the three common definitions of "Singularity" (http://yudkowsky.net/singularity/schools) then you still don't go into AGI unless you're willing to embark on the equivalent of a 30-year startup.

How many successful Internet entrepreneurs do you know? Okay, now how many successful AGI projects do you know? That tells you something about the relative difficulty here.

Basic research in the midst of scientific chaos is HARD.

Main problem with AI is that nobody really knows what it means. When something starts to be remotely useful, it is not called AI anymore :)

For me, AI's goals are very cool but most of its approaches not so much. In fact, I find most AI talks rather boring.

The only parts I really like have to do with human computation (e.g., ESP Game) and stealing ideas from nature (e.g., genetic programming, ant colony optimization, etc.):

http://video.google.com/videoplay?docid=-8246463980976635143

http://www.amazon.com/Digital-Biology-Peter-J-Bentley/dp/074...

My total out-of-butt reasoning: with low-level AI-ish things becoming commonplace (computing power getting cheaper, and devices becoming more ubiquitous; ever cheaper, more sophisticated robots, O'Reilly books on Collective Intelligence) and more data around, there'll be more demand for harder, truer-AI stuff. And if somehow there isn't, it still helps to have an understanding of the lower-level, easier stuff.

And, there's a lecture on Google video (don't have the link) of a guy going on about how there are lots of expert systems out there run by companies that don't advertise themselves.

> And, there's a lecture on Google video (don't have the link) of a guy going on about how there are lots of expert systems out there run by companies that don't advertise themselves.

It it this, by any chance? http://is.gd/otrN

I haven't seen that one myself (watching it now), but I've seen others where Larry Smith talks about exactly what you mentioned. Unfortunately nobody thought to mic him and his style of moving around makes him hard to hear on a camcorder mic.

That was a great video. I made this post with an interest primarily in the expert systems part of AI and this video solidified by interest.

I now know what I'm going to do after the internet. :)

Yeah, that's it. Too bad he didn't get more specific, give more angles of attack. Except maybe, to learn an industry and think of how to automate within it.

I think robotics is in a lull, and will be for a few years. This of those out there (pleo, roomba, packbot, kiva, anybots), and you'll find they are all "dumb". Interesting physical intelligence like BigDog isn't really a product.

It will be a few years before any intelligent robots come out. 6-10 years is a good time frame to start a robotics business.

I do AI for work, however, it is not my only task. Most of my work is development work, data wrangling and debugging. Pure AI jobs probably exist, but I like the diversity and hoping to do more research is a lost cause in most software applications since so much time is required doing development.

I work in AI as well after completing a phd in computational intelligence (a relatively recent flavor of AI). from my limited exposure, working in AI involves consulting/software engineering with different tools, in my case intelligent agents.

I took a degree in AI, and now I'm firmly convinced that AI == search. If you make something searchable in a simple, new way, that's AI. If it's grandiose but doesn't work, that's not useful.

Two exceptions:

1. Many computer vision algorithms (Canny...) are more clever than search.

2. Fourier analysis is not search.

But maybe that's too low-level to be "real" AI.

About (1), I wrote some software that finds text in an image, by searching for items and filtering with rules.

Maybe in my post, I was being too general.

I studied neural networks at University too, and maybe I don't use them because I find them hard to apply to a search task.

But all the AI module's theory (outside of neural nets) was about searching trees, etc.

AI includes heuristic search (and a great many problems can be solved with heuristic search), but it includes other things as well.

Oh well. Maybe I should just say "Here's what I learned as AI at MY University and your mileage may vary".

Shameless plug for my University: http://z.cs.utexas.edu/users/ai-lab/publications_recent.php

Shameless plug for mine, although the publications page is pretty terrible: http://www.inf.ed.ac.uk/publications/

Every time I have a conversation with someone in the department I get excited about some new direction of AI research. (I work in natural language processing myself.) AI certainly isn't dying, but from my own experiences straddling business and academia, I'd say it takes far too long for Cool Stuff(TM) to make it into the real world.

By AI, do you mean Asians and Indians?

In the Silicon Valley, yes.

Reply on news.ycombinator.com