Ask HN: So what's new in the world of A.I.? | Hacker News Reader

Ask HN: So what's new in the world of A.I.?

107 pointsdstein48 comments

AI research and related topics and startups rarely make news. Is it really such a dead field? There's never any breakthroughs announced, and there's no grandiose research projects underway that I know of. I read every day how mobile social network platforms are being funded en masse in Silicon Valley while an entire branch of computing appears to be whithering away in the grips of academia.

Are there any websites that try to track the different AI projects going on? Which books should someone fascinated in this field be reading to try stay up to date?

View on HN

Whenever there is a 'breakthrough' in AI two years onwards we call it by some different name and AI gets pushed just a little bit further in to the future again.

Plenty of stuff that is 'normal' in machine learning today would have been advanced AI twenty years ago. I don't expect this to change much, until the time that we reach the stage where you can click together a functional agent from a bunch of pre-made units that can go out in to the world and do useful work for you. And quite possibly by that time the definition of AI will have shifted to 'a machine intelligence so advanced that it can solve problems that humand can't solve' or something to that effect.

AI simply is a moving target, nobody (except for Alan Turing maybe) ever sat down and wrote a description of the border between regular programming and artificial intelligence (self improving systems) with enough clarity that we will know when we've crossed. By plenty of standards we crossed it a decade or more ago, by many standards we may never cross it at all.

Also, a lot of today's robotics (like anything from anybots, or from Andrew Ng's group at stanford) would be scary AI a few years ago.

In this sense AI can be seen to be similar to philosophy.

I'm flummoxed by this question. From what I see, AI is at the core of much of what's happening in Silicon Valley and in tech in general.

But then, I'm also using semantic web tools in building a search engine for financial news at http://Newsley.com. Right now we're indexing and automatically categorizing 500 financial news articles a day. And, I'm really looking forward to adding on recommender systems and use some machine learning with the scads of data we're collecting.

AI companies are all over the place. Any decent online ad platform is essentially a huge recommender system. Search and search quality are essentially problems that traditionally fit into a "AI" category. Have you shopped on Amazon recently? Where would they be without their recommendation system. Same with Netflix.

I just bought a $40 HD webcam that comes with software that can track points on my face in real time. It then drives a 3D puppet in sync with my facial expressions and movements. When I was studying animation 8 years ago, it took $500K worth of equipement to do motion capture. Gazehawk launched a few months ago, and are already profitable selling eye tracking tools: http://www.gazehawk.com/.

You can now buy a friggin robot to vacuum your living room for $150. The US airforce now considers commanding a wing of drones essential experience for Colonels interested in promotion. Most drone pilots in the Iraq and Afghanistan wars a pilot their aircraft from offices in Nevada. Robots are essential members of bomb and IED disposal teams. One of DailyBooth's founders was stuck in the UK because of visa problems, so last year he was attended invetsor meetings virtually by using an Anybot.

You can play virtual golf by swinging a controller in your living room if you own a Wii. The last RPG that I played was Oblivion and I was blown away by the character AI. That was years ago. I'm afraid to buy any new games because they'll suck up so much of my time.

Far from whithering away, just the opposite has happened. What used to be considered AI has now become so commonplace that we don't think twice about using it. If you don't believe me, just Google around for a while. You'll find the answer.

I know what you mean, but I think we're still at the zygote-stage of A.I.

When I go to Google and type "Where can I buy a hamburger", Google is completely useless -- but I do see 4 instances of the same Steve Martin comedy sketch from Youtube. It turns out all search engines absolutely suck for giving reasonable answers to reasonable questions. There is an enormous amount of room for innovation, and if there is innovation going on, it simply isn't making news, and it doesn't appear to be getting the kind of investment attention it should.

What about saying "Hamburger shopping basket"? That'd find you an online store selling hamburgers.

Or use Google's Froogle.

The point being made here is that we are still unable to build programs which interpret language nearly as effectively as the human brain. Our programs still have little skill in determining context in language - which will be an important step in reaching Kurzweillian levels of computing competence.

However, it should be noted that Natural Language Processing and studies in word sense disambiguation have brought us closer, people still cherish the "Dude, where's my flying car" impatience of technology.

Yes, NLP is still a tough nut to crack.

The flip side is that if you type just 'hamburger' into maps.google.com, you'll likely get a map of all the hamburger stands near your present location (based on either GPS or your IP address).

I'd argue that this is pretty damn smart, and it is way ahead of where we were 10 years ago.

I think the military drones are proof of the failure of AI research. Every single one of those devices lack AI, they must be controlled by a person. They are large remote control weapons, but not intelligent robots. The military funds plenty of research on AI but haven't actually fielded autonomous systems yet.

I think that's very good that people still control the military drones. You don't want the drone dropping a bomb where it is not supposed to just because there was some suspicious movement there. No matter how well you can program the robot and how much elaborate heuristics you can give it, at this point it seems to me that humans are infinitely more flexible in making decisions. I wouldn't allow computers make life-and-death decisions any time soon.

Unfortunately untrue.

Many military robots are purely remote controlled, but many others integrate often substantial levels of autonomy. Most UAVs provide autonomous navigation and stationkeeping, for example, and often much more; other military robots operate with no human in the loop whatsoever (eg, AUVs).

It's also interesting to consider the autonomy of fast-react weapons-in-the-loop systems such as AA/SAM sites and ABM installations (eg, Patriot), and the contribution of that autonomy to associated friendly-fire incidents. Methods normally considered AI---probabilistic filters, reasoning, and data fusion, in particular---are fundamental to such systems.

Forgive me, I failed to make this clear, but I work on such systems and I don't think they are all that smart. The code for station keeping is less interesting than the code for directions on Google Maps. (Connecting GPS points in space is pretty damn easy). Landing a UAV is pretty sweet, but I still wouldn't call it AI. The complexity is on the control side, but if you call that AI then is your cruise control in your car AI?

Well, the terminator films give reason as to why we're not fielding AI-controlled military drones.

Also, every unmanned spacecraft ever has been controlled by an AI. These drones show the success of AI research.

Well I think for one, the field of AI is massive and spans enormous amounts of research over 60-70 years. You wouldn't know about any breakthroughs unless you studied in the field.

From my brief exposure, here are some big algorithms and some applications in roughly chronological order. Definitely not exhaustive. Those in the know...feel free to correct me:

- Least squares regression (prediction- used everywhere)

- Fishers Discriminant (classification tasks)

- Perceptron networks (classification tasks)

- Markov Models/Hidden Markov Models (Speech/Handwriting recognition)

* Machine Learning community develops out from AI *

- Support Vector Machines (Image recognition/ classification)

- Expectation Maximization (prediction)

- Relevance vector machines (classification / prediction)

- Gaussian processes

- Predictive sampling

My prediction (heh pun intended) is that you see enormous changes in the field when processing by GPU's becomes much more available. There are some algorithms that are simply difficult to research because labs don't have access to fast enough machines. Also, there is beginning to be some effort to port these algorithms to a Map/Reduce framework so they can be run at scale (check out Apache Mahout). Lastly, I'm slightly biased towards machine learning as that's where I chose to do my grad research. I'm not sure what problem domains are under AI or ML or Statistics...I tend to clump them all together.

Gauss invented least squares. Before 1800 IIRC. Also, I had no idea svm predated em. Cool.

As for the gpus... I'm dubious. Most optimization problems I'm aware of are poorly suited for gpu style parallelism. More's the pity.

Genetic algorithms work pretty well for obvious reasons.

An obvious choice would be ridiculously large Neural Networks. However, any problem where you start with a random seed and converge on a solution benefits from starting at several different random seeds to avoid local minima. Also search algorithms (chess programs etc) often paralyze vary well.

Monte Carlo AIXI: http://www.vetta.org/2009/09/monte-carlo-aixi/. It's an approximation of the optimal-but-uncomputable AIXI agent. Apparently it learned to play a Pacman variant without being told the rules in advance.

In today's context, most work in AI is statistics-based. So if you are talking about conventional AI (logic-based for example), yep it would be relatively harder to find.

Lets not forget though that machine learning and data mining and their applications: natural language processing, a lot of computer vision, recommender systems, financial predictions, fraud monitoring, some types of games, and innumerable others are huge advancements made towards the same goal that AI was after, albeit with different means (statistics) than originally attacked with.

Statistical machine learning is AI. And it is bigger than ever before.

Large-scale data-mining, predictions, and machine learning are not new per se, but have rapidly expanded in terms of available practical techniques, and actual use of those techniques in applications. Palantir is one major startup based around that, and there's a bunch of other less-high-profile stuff moving AI into marketing (behavior prediction, etc.). There's also quantitative finance. Less businessy, there's a bunch of research on applying machine learning and datamining to huge datasets like NASA's sky surveys, or biology or physics data repositories.

To add to the many good responses here so far: AI is such an overloaded term with many negative connotations that it's no longer used. It's also now specialized into separate components, as researchers realize that it's not "cheating" to study e.g. vision separately from natural language processing. Even humans cheat, by having separate portions of the brain specialize in different parts of "AI".

So some of the major AI fields are now known as:

  * machine learning - learning from data, aka applied statistics, aka how to "learn"
  * computer/machine vision - how to "see"
  * natural language processing - how to "read" and "write" and "translate"
  * speech recognition - how to "hear"
  * robotics - how to "do"
  * affective computing - how to "feel" and "act"

There are others as well, but as you can see from this simplistic breakdown, the specialities loosely mirror those in biology/medicine/neuroscience, and indeed there are also researchers who straddle the boundaries between the natural side and the computation side of things.

As a vision researcher, I can tell you that there's a huge amount of work being done, and progress being made, in our field -- both in academia and in industry. While it often doesn't make the news, we consider this a feature, not a bug.

For following progress in these fields, I have two comments:

1. There's no good resource I know of to follow all the subfields of AI. Instead, there are different sources for each field.

2. While others have recommended journals and conferences, I think it can be tough to read them if you're not already immersed in the field. So instead, I'd recommend starting with the wikipedia pages for each field, seeing the general list of topics, and then finding the appropriate papers if you're really interested. A good way to find important papers is by looking on google scholar for papers with lots of citations (insert usual disclaimers here about citations != quality of work, etc.)

I can get you started in computer vision with two very influential papers in the last decade that have also had a huge impact on industry:

P. Viola and M. Jones - Robust Real-time Object Detection http://research.microsoft.com/en-us/um/people/viola/Pubs/Det...

This paper revolutionized face detection, and is the basis for automatic face detection in most consumer cameras.

D. Lowe. Distinctive image features from scale-invariant keypoints http://www.cs.ubc.ca/~lowe/papers/ijcv04.pdf

This paper was the culmination of many years of work on detecting repeatable features in images and representing them in a consistently-findable way and is the basis for numerous object recognition algorithms, as well as those for stitching multiple photos together into panoramas and Microsoft's "PhotoSynth".

I don't think it's true that it's no longer used; I certainly still consider my primary research field to be "AI", and the main journals and conferences I publish in have "AI" in the title. I do agree that some subfields have split off more, especially machine-learning, robotics, and vision, but there's plenty of stuff that still goes on under the rubric of AI. Even the more split-off fields are still very present in big-tent AI conferences and journals, especially when presenting work on integrated systems.

Despite being mostly known as an ML researcher, Leslie Pack Kaelbling's AAAI-10 keynote slides make something of an argument for why AI is important as well: http://people.csail.mit.edu/lpk/AAAI10LPK.pdf

The skills and techniques required are very very different between the fields though; my PhD was in natural language processing but I don't have the first clue about machine vision, for example.

Hell, I was working on parsing biomedical text -- meaning I don't really have much of a clue about speech processing, machine translation, sentiment analysis, question answering or most of the other subfields of linguistic computing.

Of course there's plenty of room for cross-fertilization, but that's true across comp sci as a whole. Thinking of all of these things as being part of some coherent topic called AI is more of a historical legacy than a useful category.

I guess taking a mostly integrated-systems, applications-focused view, I see AI as a useful organizing concept for what to build, what the issues are in building it, etc. For example, I'm not sure what field Alan Newell's "knowledge level" talk would go in if not into AI.

I do agree there are lots of areas of research that are maybe more "algorithms" than "AI" (e.g. improving SAT-solving), but I disagree that that sort of specialization is the only way to research. From my perspective, those areas of algorithms provide the raw-material research that can be used to build AI systems. Even building AI systems is often specialized to some extent as well, but I think it's useful to have a semi-coherent body of knowledge and shared community around "AI" when doing so, rather than just the domain-specific algorithms and approaches.

... and as someone whose research involves both SAT and statistical learning, I'd agree: putting a variety of specialties under the AI umbrella can make it easier to connect different areas in useful ways.

Seconding the recommendation for David Lowe's paper. It's easy to follow and seriously impacted image detection. If you have MATLAB you can also run his demo: http://www.cs.ubc.ca/~lowe/keypoints/

It's very interesting to see.

As others have pointed out, a lot of things that "work" stop being considered AI.

But here are a few recent advances off the top of my head:

The DARPA Grand Challenge. Robot cars that drive themselves. http://www.darpa.mil/grandchallenge/index.asp http://www.youtube.com/watch?v=RY93kr8PaC4

Google's Machine Translation system. Far from perfect, but it can usually do something reasonable between over 40 different pairs of languages. They added Latin on Thursday.

CMU researchers have developed a classifier that can look at your MRI scans and tell what you're thinking. (Still in its infancy, but still.) http://www.computerworld.com/s/article/346917/The_Grill_Tom_...

People working in my research lab are working to automatically reconstruct words from the ProtoAustronesian language using just the modern words. Our results are typically extraordinarily close to what historical linguists have done by hand.

People in a neighboring lab are working on a system that can automatically distinguish between a nuclear detonation and an earthquake as part of the Test Ban Treaty Organization. They're making good progress.

Take a look at PhotoSynth: http://photosynth.net/ which can automatically reconstruct 3D scenes from disparate photos taken from different cameras with no prior information needed to stitch them together. http://grail.cs.washington.edu/rome/

The face detection that is now standard on many cameras was one of the great problems in computer vision.

The field is most certainly not dead. Mass media (movies, books, etc) over the last 60 years or so has presented a vary narrow vision of "A.I"; a vision of what are more or less people that happen to be in synthetic versions of the old flesh and bone varieties, that are normally repressed and enslaved or trying to kill the carbon-bags for some faulty reason or another.

That vision has basically made the term Artificial Intelligence become yet another meaningless marketing term.

The concept of A.I. is like the concept of organic chemistry. There are vast sub-disciplines that are loosely collected under the term "organic chemistry" because they more or less all involve carbon atoms at some point. A.I is like this as well, because there are many, many sub-disciplines and they all involve the concept of systems that are capable of "learning".

So, my first recommendation is to just forget any preconceptions of "Artificial Intelligence" because its not something that really exists, just like Brawndo the Thirst Mutilator does not have what cows crave.

What does exist is this vast field of scientific, practical, and commercial activity involving learning systems. These systems are typically closely tied to computation, aka hardware or software, but not all. Some work is done in biology and some in organic chemistry (interesting cross connection there).

Another preconception to forget is that "A.I" died in the 80's and is irrelevant today. What nearly all people who are even loosely familiar with the term, beyond killer robots from the Future or lonely little robot boys with mommy issues, is that A.I. has no part in their lives. The truth is that it has been all around us for many years.

The timely delivery of your new shoes from Zappos or buying oranges in the winter: shipping. Tens of thousands of planes in the air at any moment: transportation. The computer chips that are in your computer or handheld that you are reading this on: computing (simulated annealing, fascinating!). Millisecond stock trades: the economy. Automated stellar classification: astronomy. The list goes on.

So once you put aside your preconceptions for a bit, start asking what kinds of fields you're really interested and then looking at how learning systems are influencing them. Sometimes you will be surprised to learn that what might be classically called "A.I" has utterly invaded your favorite field and that the people who work in it are entirely unaware that the learning systems they're using might even be considered "A.I".

Numenta (http://www.numenta.com) is working on simulating how the brain (the neocortex in particular) recognizes patterns. It's an exciting recent field of research, and has little to do with neural networks, or other supposedly brain-inspired methods that don't actually bear much resemblance to neural processes.

This TED talk by co-founder Jeff Hawkins (of Palm fame) gives an overview of what they are doing.

http://www.ted.com/talks/jeff_hawkins_on_how_brain_science_w...

People have already explained the fact that the field now goes under a number of different names, and that it is very active, so I won't rehash this. A few pointers to things where you can keep track that are different from what others suggested so far.

1. Foundations and Trends in Machine Learning -- this is a journal aimed at publishing a very small number of well-written survey papers on various trends in ML. This is easier to follow than an entire conference (much lower traffic, higher signal/noise), and should be readable for a wider audience (assuming they are math-inclined).

2. Conferences like Algorithms for Modern Massive Datasets are practically-oriented, well attended by a lot of industry, and involve a lot of AI: http://www.stanford.edu/group/mmds/. Look through the speakers and topics. This is one example, there are others.

3. A lot of important tech companies have teams that do AI and AI-type things, at least using the modern definition of AI (Google, Facebook, Twitter, LinkedIn, Netflix, Amazon, Microsoft, eBay, even Apple with its Siri acquisition; there are others). This is not to mention people using this stuff in other areas, like finance and bioinformatics. These groups sometimes talk about what they're working on, so you can check this out.

On MetaOptimize, we were discussing the most influential ideas from 95-05:

http://metaoptimize.com/qa/questions/867/most-influential-id...

And just to clarify: Most of the contributions have been in machine learning. Machines can learn a superset of AI (assuming that machines can learn AI) because there are a handful of things a machine can learn that a human can't.

> AI research and related topics and startups rarely make news.

Research of any sort rarely makes news. The exceptions involve universities with very zealous, hard-working P.R. people (who pretty much always overstate the importance and impact of the work they cover).

As for startups, they, and other businesses, are there to make money. And there just isn't much money to be made in A.I. (yet?). The only commercially viable idea to come out of A.I., that I know of, is the expert system. And that turned out to be a pretty small market. [EDIT: Seems I was way off here. See replies, and my apologies to all the hard-working A.I. researchers out there.]

> Is it really such a dead field?

Well, not dead, but it certainly is not what it once was. 50 years ago, many knowledgeable people thought that machines with human-level capabilities were just around the corner. Today it's just Ray Kurzweil. 50 years of rather dismal results tends to dampen the enthusiasm of prospective researchers.

> There's never any breakthroughs announced, ....

"Breakthrough" is a word used by P.R. people. Real researchers know that knowledge is almost always advanced in small steps, building on previous work. When you read about a "breakthrough"--in any field--be skeptical.

> ... and there's no grandiose research projects underway that I know of.

Almost certainly not. This is because the U.S. Congress, and similar organizations elsewhere, have not made A.I. a funding priority. No grant $$$ means no big research projects at universities. (And no obvious commercial possibilities means no big research projects at corporations.)

But there are certainly plenty of little projects.

> Are there any websites that try to track the different AI projects going on?

That would be difficult, since, when a researcher starts a project, he generally does not announce the fact to the world. Research projects often have no official status. It's just a few people--or one person--playing around with some ideas. If they come up with something interesting, then they publish a paper about it, but that is only after the fact. So it would be hard to gather the information for such a website.

Generally, the way to track current research in a field is to look at the relevant journals. Unfortunately, this can be tricky without expertise in the field. In any case, here are a few links to A.I. journals, to get you started:

Journal of Artificial Intelligence Research http://www.jair.org/

Journal of Machine Learning Research http://jmlr.csail.mit.edu/

Journal of Intelligent Systems http://www.jisjournal.org/

Another possibility is to find an online discussion group about A.I. research. However, I can't help you there.

> The only commercially viable idea to come out of A.I., that I know of, is the expert system. And that turned out to be a pretty small market.

Oh come on, you're baiting AI people (like me) here. :P

* Spam Filters (Bayesian networks)

* Video Game AI (game trees, fuzzy logic, etc)

* Natural Language Processing (Markov networks)

* Computational Finance (pattern recognition)

The list goes on, but the thing about AI is this: if an algorithm is successful, it quickly becomes a part of the industry or its own field and is forgotten.

Thanks for the correction.

Amen. Stuff that doesn't work is "AI". Stuff that does work is "math".

An algorithm to predict which humans are trustworthy enough to keep their promises to their creditors, which murders the performance of human beings, enabling billions of dollars of commerce which was previously economically unfeasible? AI. FICO? Oh, that is just some number crunching on a big source of data. Big deal. Come back when you get a chatbot working.

When most people think AI they think systems being able to make decisions like humans, ones which use logic but we can't always deduce why they have been made.

Once a deterministic route to a decision is shown it does now feel like math and a lot less like the grand ideas they had in their minds.

Hofstadter said in GEB that "intelligence" is almost always defined implicitly as "whatever we currently think computers can't do," and that goalpost gets moved with astonishing regularity. The fact that a machine can do something means that it must not take intelligence, right? Thus, computers are able to do more and more that would have been thought impossible just 5 years ago, and we still convince ourselves that A.I. lies far beyond the horizon.

Wow!

Way to undermine the contributions of the AI community.

Machine Learning and Natural Language Processing have already made great sides.

Search, recommendation engines and automated translations are all AI.

Disease prediction/diagnosis has been pretty hot for a while now. Market for predictions in Analytics is picking up rapidly and there's a lot of focus there.

Reinforcement Learning is playing huge role in modern Robotics.

Sentiment analysis i.e. judging sentiments of authors from text is pretty hot.

DB community has long been working with AI community in developing semantic storage systems. Work on adaptive systems is also seeing a comeback.

I really doubt any piece of major equipment today is built without AI.

The part which I did like about your response is that reading journals is a great way to learn more about AI stuff.

> Way to undermine the contributions of the AI community.

Apparently.

I think I know a fair amount about research in general, but not, it appears, much about the current state of A.I. research.

The truth is that a lot of the sites that you use on a regular basis -- Google, Facebook, Amazon, eBay, last.fm, your bank, your webmail provider, etc. etc. etc. -- all rely on different kinds of AI for parts of their functionality.

You are exposed to and using AI all the time.

If you want AI literature treating human-level intelligence, try searching for "artificial general intelligence".

Also, here are some links to the proceedings of recent AGI conferences:

http://www.atlantis-press.com/publications/aisr/AGI-09/

http://www.atlantis-press.com/publications/aisr/AGI-10/

There is the Singularity Institute for Artificial Intelligence -- http://singinst.org/

They are tackling the problem of how to make AI "friendly" -- http://wiki.lesswrong.com/wiki/Friendly_artificial_intellige...

Our ideas are definitely quite a bit different; track THIS: http://artificial-intuition.com (theory/motivational site) http://videos.syntience.com (videos - most recent results) http://monicasmind.com (blog) http://syntience.com (corporate. Investor inquiries welcome) http://ai-meetup.org (something to do in Silicon Valley) http://www.youtube.com/watch?v=REzrYWOzhWc (bonus robot)

The "dead field" you speak of is deterministic A.I. research. Statistical methods (read: machine learning) are doing extremely well these days, and only get more powerful the more data we get our hands on.

I'm sure my karma rating will take a hit, but my first thought was -- "Someone developed a program to replace Mike Arrignton and his ability to insult interviewees?"

But in all seriousness, the links to journals are your best resource for specific topic areas. I've spent a lot of time in data modeling, computer vision, etc. It is a lot about providing specific models (with some minor adaptation / dynamism in the algorithm) to provide for "learning" / "adjustment".

So start with areas that are directly related to your interests and try and iterate/generalize from there.

... I like the fact that 15 places above this post, currently, is this one:

http://news.ycombinator.com/item?id=1754357

One recent development in the field of artificial neural networks is "Deep Learning": http://www.youtube.com/watch?v=VdIURAu1-aU&feature=playe...!

Apple bought Siri, a spinoff from SRI's AI division, earlier this year. That was a pretty big deal.

how bout amazon's recommendation systems? or ad placements? or google search? google what did you mean? ocr? algorithmic trading?

what field are they in if not AI? sure, the techniques are heterogeneous and there seems to be no one unifying theory but the field is a bit like that.

Reply on news.ycombinator.com