Vicarious Systems Says Its Artificial Intelligence Is The Real Deal | Hacker News Reader

Vicarious Systems Says Its Artificial Intelligence Is The Real Deal

blogs.wsj.com

48 pointschurp36 comments

I cannot understand why it is that so many obviously very intelligent people decide that we need another computer vision-based startup. Because the unfortunate truth is that computer vision (right now) doesn't work.

Let me qualify that. From the academic / research point of view, there have been a collection of real successes in computer vision in, say, the last ten years. But my sense is that what counts as a research success is a long way from what counts as a practical business success.

For example, the best generic object detector at the moment is probably Felzenszwalb's using deformable parts-based models[1]. And it's just not that good. On the latest PASCAL object detection challenge, you'll see that its mean precision is only ~30%.

Scott Brown, the interviewee, sets Vicarious apart by highlighting the fact that their system will be neurobiologically inspired. But the idea of learning hierarchical systems that mimic the brain's visual processing system is hardly new, and the jury is still out on whether these systems can do better than the "hand-coded" systems like Felzenszwalb's. As a random example, see [2].

Like.com showed you can build a business that uses computer vision in some way. But as Brown snarks, they "use a big bag of different heuristics to figure out the image." For the time being, that seems to be the only way to get computer vision to work in practice.

That all said, I wish them luck.

[1] http://people.cs.uchicago.edu/~pff/latent/

[2] http://www.cs.stanford.edu/people/ang//papers/nips07-sparsed...

Your argument against computer vision startups is that there isn't a viable computer vision solution at this point?

Well, his argument is that well-funded, very intelligent people are trying like hell at computer vision, and not succeeding. That's not a good sign - you'd prefer that your space has been hitherto overlooked by smart people with lots of money.

I think he's arguing that computer vision is a research subject - most startups are doing known things (in the sense of "this has been successfully done before") or at most development ("this has been successfully done before - in the lab").

> I cannot understand why it is that so many obviously very intelligent people decide that we need another computer vision-based startup. Because the unfortunate truth is that computer vision (right now) doesn't work.

This seems like a really good reason to create another computer vision-based startup.

No, it seems like a very good reason to take useful/promising but improperly commercialized research and turn it into a product. A startup rarely has enough runway to do the scientific research needed to solve a problem like this.

It is certainly rare, but Numenta has been doing it for the past 6 years, and for several years before that at the Redwood Neuroscience Institute from which it spun off. In doing so, Numenta undoubtedly stands on the foundation of significant progress in academia, but still has to do a fair bit of what one might call "research engineering."

Basic R&D is a cost that successful businesses--especially small ones--tend to externalize. Microsoft does a bit, Bell Labs used to do more; but you just don't start a FTL spaceship company before basic research has established a coherent theory of FTL travel.

And yet facial-recognition is now freely available to consumers (Picasa, Facebook etc), our phones have blink detection, 3D motion detection and tracking is available to consumers for ~$100 (Kinect).

I'm not familiar with the PASCAL object detection challenge, but I just had a quick look. It's hard - if I understand it correctly, classifiers had to categorize photos into containing 5 types of objects form the 1000 leaf nodes of http://www.image-net.org/challenges/LSVRC/2010/browse-synset.... (Based on the description from http://www.image-net.org/challenges/LSVRC/2010/pascal_ilsvrc...). I'm having trouble understanding the scoring scheme (how is flat cost calculated?), but based on this I'm quite impressed.

I'm human (yes, I swear it's true), and I couldn't classify things like different breeds of poodle: http://www.image-net.org/synset?wnid=n02113712

There are many different actual tasks that technically are PASCAL challenges, but when people say "PASCAL VOC challenge" (Visual Object Classes), they typically mean either the _classification_ or _detection_ challenge:

Classification: For each of the twenty classes, predicting presence/absence of an example of that class in the test image.

Detection: Predicting the bounding box and label of each object from the twenty target classes in the test image.

Neither uses the full ImageNet data set. Instead, it's images from 20 classes of object, like shown here: http://pascallin.ecs.soton.ac.uk/challenges/VOC/voc2010/exam...

Here are the results: http://pascallin.ecs.soton.ac.uk/challenges/VOC/voc2010/resu...

I find the results quite impressive - especially for classification - 90%+ precision for detecting people in photos seems like a good result.

As an AI grad student, this kind of sensationalism is somewhere between a minor irritation and a serious threat. AI always has had a severe problem with over-promising and under-delivering, and I'm of the humble opinion that until you're actually shipping the most awesome thing in the world you should keep your mouth shut. If the first thing people associate "AI research" with is "disappointment", that hurts everybody (particularly, NSF funding).

"Brain-based" AI should stay in the dark ages. Optimization-based AI is the present and the future.

(That said, if you want to talk about your sweet computer vision system that's "coming soon", go right ahead. Just don't call it AI.)

Does Numenta fall under "Brain-based?"

Numenta does fall under brain based. I'm not sure what Vicarious are working on, but recently Numenta transitioned to radically more biological algorithms. It would be interesting to compare the two if Vicarious comes out with more detailed information about their algorithms.

I would say "Brain-inspired". As I see it, Numenta's model (as of about a year ago) is based on (1) the hierarchal organization of neurons, (2) the presence of feedback loops in neural architectures and (3) the importance of temporal processing even for static scenes. This doesn't include any intracellular details nor any of the larger and/or specialized brain structures.

> AI always has had a severe problem with over-promising and under-delivering

Is this because the AI researchers truly over-promise, or because media/laypeople take a concept or statement and run with it?

Traditionally, it was the AI researchers over-promising. See the Wikipedia article on the AI Winter: http://en.wikipedia.org/wiki/AI_winter

There was something of a collective and large-scale underestimation of how hard AI would be. Why that would be the case is interesting, especially since it persisted for some decades, across multiple fields filled with smart people. From probably the 1880s to, say, the 1970s, there seemed to be this widespread view that high-level AI was just around the corner, and mainly depended on some inevitable technical progress (faster computers and more memory, plus a bit of algorithms work).

There wasn't even really much debate, in either CS or philosophy or engineering, over whether computers would be able to do "routine" tasks like accurate object recognition, mathematics, playing chess, etc., in the near future. The biggest controversy was over whether computers could ever be "truly" intelligent and creative, e.g. whether computers would also replace Beethoven in addition to mathematicians, or whether they'd be forever limited to just being very capable automatons. Somehow everyone missed that even making them the "lesser" kind of intelligent, so they can walk around, recognize objects, translate languages, etc., would turn out to be pretty hard.

"Brain-based" AI should stay in the dark ages. Optimization-based AI is the present and the future.

Humans can see. Computer vision systems suck. There's a perfectly good one in our brains. Why not try to understand what already works?

Contrary to what most would believe, brain-based computer vision has made a lot of progress in the past 20 years. Some might think there is a fundamental flaw in the "brain-based" approach given past failures, but that ignores that fact that those failures very likely happened due to a poor understanding of the brain at the time.

The work in brain-based computer vision however has been mostly academic. Brain-based computer vision startups are even more recent, and I think it's exciting to see the startup approach to solving what has been mostly an academic problem. In a startup, the engineering mindset, quick iteration, as well as a lack of concern for publishing and other forces at play in academia could produce very different results.

I do agree that the 5 year promise is extreme, but I think we need time to see how this relatively new mode of work (both in terms of the technical approach, and the process of implementation in a startup) will play out before we call it a failure.

Full Disclosure: I was an intern at Numenta last summer.

Is Numenta's approach really that informed by findings regarding actual brain function? It's been a while and I don't remember most of Hawkins' model, but I don't feel that one needs to consult any actual neuroscientific results to use the general concepts of hierarchical design or top-down processing, which seem to capture the basic idea of his work.

This whole neuro-A.I. fad began with artificial neural networks, which had nothing to do with brains, and still hasn't died.

Numenta's most recent algorithms are actually very strongly neurobiological. If you have looked at earlier work, you should check out the most recent white paper from a couple of months ago, which detail more than year of recent efforts in that direction.

http://www.numenta.com/htm-overview/education/HTM_CorticalLe...

There is also a recent talk by Jeff Hawkins from a few months ago on the same subject.

http://www.archive.org/details/Redwood_Center_2010_12_02_vs2...

You are correct that neural networks had almost nothing to do with brains. Numenta's new cortical learning algorithms, on the other hand, are very closely modeled on the structure and function of the neocortex.

> this kind of sensationalism is somewhere between a minor irritation and a serious threat

A serious threat to what?

Continued funding, I'd guess. It's the same sort of talk that triggered the AI Winter.

Speaking from a researcher (both academic/industry) in vision for almost 10 years, I am afraid that they founders have quite underestimated the difficulty of the problem. Even a dog's visual system is very advanced, if you consider it from the big picture in evolution of visual sensory system in animals. So in the interview "if you can make a vision system that’s just as good as a dog" is in some sense analogous to saying "if you can simulate what's produced from 90% of visual evolution over these million years", which is clearly, a bit over-optimistic as a starting goal.

That being said, wish them luck. It's a worthy try afterall.

More accurately: Vicarious says it is trying to make its vision system the real deal. Smaller problem, and (at least according to the article) they haven't solved it yet.

Of course, smaller and small are different things - this is still a very hard thing to do. Hope they succeed.

Does this mean the AI winter is over?

Depends what you mean. "Good old fashioned AI" (explicit symbolic representations and rule-based inference) hasn't made a comeback yet. Even computational linguistics has largely shifted towards statistical methods based on crunching large amounts of data. So, AI is a lot livelier now than it was in the bad periods, but it's a different kind of AI than went cold in the late 1980s.

I'm not an AI student myself (I do CS, but with an interest in AI and probabilistic algorithms) and since I'm from a younger generation most of the "new" AI stuff like neural networks, evolutionary computing, etc. wasn't that new any more when it got taught. But "classical AI" (as in symbolic AI and rule-based systems) always struck me as rather dumb and inefficient after being shown the power of probabilistic systems like EC and neural networks. So in that sense I don't think classical AI will never make a comeback, and I'm convinced this is probably a good thing.

I mean, the only example we have of solving extremely complex problems is nature, and nature just doesn't work with symbolic system and rule-based inference. It uses probabilistic systems which interact with each other in feedback loops.

I, for one, welcome our new non-deterministic overlords.

Interestingly, one of the cofounders is Dileep George, previously the CTO and cofounder of Numenta.

From the article

"if you can make a vision system that’s just as good as a dog..."

Not quite my idea of "The Real Deal". And that's within a 5-year plan.

Dogs' visual systems are pretty sophisticated. Trying to mimic one of those first allows one to somewhat simplify things while getting a lot of insight into the human visual system which operates on basically the same principles.

NFLT implies AI is logically impossible. Good way to short the stock market.

Considering NFLT isn't on the first page of google, would you mind expanding that acronym?

If the parent means the No Free Lunch Theorem, then the point is mistaken. That theorem says that you can't improve the performance of a classifier for some objective function, without making it worse on another -- in other words, all algorithms have an identical mean performance when averaged over all possible objective functions.

The reason this doesn't mean that human-level AI is impossible is that we too are designed (well, evolved by natural selection) to perform well for a particular objective function: one in which say, the standard laws of physics/optics apply. Optical illusions illustrate that our performance on this objective function is not perfect.

Moreover, you can see a human being's performance on a different objective function by, for example, trying to recognize objects in pictures which have been scrambled according to some predefined method (e.g. shuffle the pixels but use the same random seed each time). Each scene will still convey the same amount of information about the objects in it, but it'll be pretty tricky to recognize the objects.

"The reason this doesn't mean that human-level AI is impossible is that we too are designed (well, evolved by natural selection) to perform well for a particular objective function: one in which say, the standard laws of physics/optics apply."

That's an assumption no one has ever given the slightest shred of evidence for. I remain highly skeptical.

It does no such thing, it is very likely that AI will be made up of a bunch of specialized interacting subsystems. As for No Free Lunch Theorem. See: Coevolutionary Free Lunch. Which by the way, is actually more akin to biological evolution than coevolution. http://cs.calstatela.edu/wiki/images/1/15/Wolpert-Coevolutio...

Reply on news.ycombinator.com