Creating a marketplace for artificial intelligence
7 comments
A bunch of classifiers is not an artificial general intelligence. A human-level mind (or beyond) requires a systematic cognitive architecture, it's not going to emerge out of a heterogenous, quasi-random soup of mind-components. This is a naive theory of cognitive science, I would argue.
For a careful, systematic design for an advanced AI system, see http://opencog.org/wiki/OpenCogPrime, which is associated with the open-source AI project opencog.org
Anyway, I'm probably one of the biggest optimists in the AI research community, but in my view the idea in this post represents "way over the top" optimism based on a naive view of how mind works.
Given a mature version of an AI system built according to an overall AI architecture like OpenCog, then maybe people writing little "mind modules" as you suggest could make sense. But the Internet just is not a "mind operating system."
I wrote about what would need to be done to turn the net into more of a mind-OS in my 2001 book "Creating Internet Intelligence" (Plenum Press). At the high level I don't think this is a bad idea. But again, brains are not just soups of heterogenous processes -- the right high-level cognitive architecture is required.
-- Ben Goertzel novamente.net agiri.org singinst.org goertzel.org
For a careful, systematic design for an advanced AI system, see http://opencog.org/wiki/OpenCogPrime, which is associated with the open-source AI project opencog.org
Anyway, I'm probably one of the biggest optimists in the AI research community, but in my view the idea in this post represents "way over the top" optimism based on a naive view of how mind works.
Given a mature version of an AI system built according to an overall AI architecture like OpenCog, then maybe people writing little "mind modules" as you suggest could make sense. But the Internet just is not a "mind operating system."
I wrote about what would need to be done to turn the net into more of a mind-OS in my 2001 book "Creating Internet Intelligence" (Plenum Press). At the high level I don't think this is a bad idea. But again, brains are not just soups of heterogenous processes -- the right high-level cognitive architecture is required.
-- Ben Goertzel novamente.net agiri.org singinst.org goertzel.org
I'm trying to draw a parallel between how content on the web is being created and how an artificial mind might be created. If you sat down at tried to engineer a giant encyclopedia within one company/community, you would end up with britannica (within one company), and wikipedia (within a community of enthusiasts), However, the web at large is much larger than both of these because of the economic incentives for people to create articles. Many people try and fail (in effect working for free), but the progress is rapid, as in evolution.
In the same way, incentivizing people to produce classifiers would have the similar effect of rapid progress.
If you think of intelligence as being the ability to predict accurately, having a giant web of classifiers that predict accurately could be construed as a form of artificial general intelligence, or a human type artificial intelligence. Having an enough data about the world, because you have created many classifiers that can recognize the semantic events in video and speech, would allow you to make all the same types of recognitions and predictions that a human would make.
If you think of intelligence as being the ability to predict accurately, having a giant web of classifiers that predict accurately could be construed as a form of artificial general intelligence, or a human type artificial intelligence. Having an enough data about the world, because you have created many classifiers that can recognize the semantic events in video and speech, would allow you to make all the same types of recognitions and predictions that a human would make.
I disagree that creating a useful artificial intelligence product requires tremendous resources. When search engines first came into existence, a few machines for a crawl and a few machines to do basic posting-list serving were enough to have a useful product. Now it's tougher but Cuil with a dozen engineers has done a reasonable job from scratch.
The real problem is to find an application for artificial-intelligence-style programming that is actually useful. What good is classifying what's in a picture? If you could sketch out a useful product then the right small group of engineers could build it.
The real problem is to find an application for artificial-intelligence-style programming that is actually useful. What good is classifying what's in a picture? If you could sketch out a useful product then the right small group of engineers could build it.
Surveillance technology is one such application already in use (location of faces and people in images), video analytics on google.
A much larger market is robotics. (Factories use robots that are custom programmed with vision for each application and don't really do recongition in the sense of object recognition in an arbitray scene). Recognizing common household objects would be useful for a company building home robots (of which there are a few) right now.
There is a recent startup doing image recognition using humans, on photos at flickr.
Search engines relied on letting a single algorithm go out and process data. There isn't yet an artificial intelligence algorithm that can just go out and do what it needs (unsupervised learning algorithm that works well). The currently available algorithms (supervised learning) require labelled data, and since there are millions of categories of things (that we as humans can percieve and so are useful to recognize), it would require at least a billion labellings of the data (1000 data labellings for each category).
The trouble with companies and groups engineers is that you would have to fund them, organize etc. and only a few groups would form (there are only an handful of video analytics companies probably, all of them probably do recognition of a handful of the same things). Besides, the system advocated above is pretty simple, hardly any engineering in the sense of specialized technical skill is required. You would just collect and feed data into the classifiers. Others would combine data from classifers into yet more classifiers.
There is a recent startup doing image recognition using humans, on photos at flickr.
Search engines relied on letting a single algorithm go out and process data. There isn't yet an artificial intelligence algorithm that can just go out and do what it needs (unsupervised learning algorithm that works well). The currently available algorithms (supervised learning) require labelled data, and since there are millions of categories of things (that we as humans can percieve and so are useful to recognize), it would require at least a billion labellings of the data (1000 data labellings for each category).
The trouble with companies and groups engineers is that you would have to fund them, organize etc. and only a few groups would form (there are only an handful of video analytics companies probably, all of them probably do recognition of a handful of the same things). Besides, the system advocated above is pretty simple, hardly any engineering in the sense of specialized technical skill is required. You would just collect and feed data into the classifiers. Others would combine data from classifers into yet more classifiers.
Another application is in mobile phones that recognize objects, already being done in japan for certain objects in stores. There was also an application developed for the android platform that did some type of recognization of outdoor environments.
Sounds pretty cool. So why don't you make a mobile phone app that recognizes some sort of useful object to recognize? That doesn't sound like it's too big of a project for a startup.
Even if a large company was to throw a huge amount of resources at this, they would still probably have a team of a few people make a demo first. They go through the same process as a startup in a sense, first having a few people trial and demo and if it seems good then give them more money and/or engineers.
Even if a large company was to throw a huge amount of resources at this, they would still probably have a team of a few people make a demo first. They go through the same process as a startup in a sense, first having a few people trial and demo and if it seems good then give them more money and/or engineers.
My point is about the "democratization" of classifer creation. I personally am not a programmer and don't have the talent to create a startup, and there are millions more people like me. However, if there were a place where i could feed some data and have a classifier hosted somewhere I could participate in this.
It's the same way blogs allowed many more people to start creating content on the web, whereas previously it required creating and hosting a website. Or how google app engine, might increase the creeation of web applications because previously a lot of people didn't have knowledge of administering and scaling unix/dbms based applications.
A large company would seriously have a look at roi, whereas many small entrepreneurs just have a go at it. Britannica (large company) is smaller than wikipedia (community of individuals), which is again smaller than the content on the web at large (community of incentivized small entrepreneurs).
It's the same way blogs allowed many more people to start creating content on the web, whereas previously it required creating and hosting a website. Or how google app engine, might increase the creeation of web applications because previously a lot of people didn't have knowledge of administering and scaling unix/dbms based applications.
A large company would seriously have a look at roi, whereas many small entrepreneurs just have a go at it. Britannica (large company) is smaller than wikipedia (community of individuals), which is again smaller than the content on the web at large (community of incentivized small entrepreneurs).
I think you're defining artificial intelligence too narrowly.
If you branch beyond the classic uses (speech, vision, text-classification) to some of the things that have been more recently branded 'Collective Intelligence', there's really a large (and quickly growing) field out there.
Bayesian filtering is standard in anti-spam solutions, recommender systems and price modelers are huge in online shopping and rental verticals, and social networks are starting to use things like clustering to connect people.
AI is making some huge advances (especially in the past 2-3 years as the 'open data' culture gains steam and the tech catches up) and I think there's a lot of room for startups in the space serving niche purposes like custom recommenders, especially when you throw data-visualization into the mix.
Examples off the top of my head include Farecast, the airline backend at ITA Software, and some of the upcoming work in the automated rss recommender area.
If you branch beyond the classic uses (speech, vision, text-classification) to some of the things that have been more recently branded 'Collective Intelligence', there's really a large (and quickly growing) field out there.
Bayesian filtering is standard in anti-spam solutions, recommender systems and price modelers are huge in online shopping and rental verticals, and social networks are starting to use things like clustering to connect people.
AI is making some huge advances (especially in the past 2-3 years as the 'open data' culture gains steam and the tech catches up) and I think there's a lot of room for startups in the space serving niche purposes like custom recommenders, especially when you throw data-visualization into the mix.
Examples off the top of my head include Farecast, the airline backend at ITA Software, and some of the upcoming work in the automated rss recommender area.
True, they are applications of machine learning techniques and can be considered "narrow ai".
What I was suggesting is more along the lines of something that can do the tasks of children, see and recongize most things, and understand language. I think one way of achieving that is by "bootstrapping" from understanding of video, If an understanding of what a human does in the video (standing, sitting, limb positions, height, facial expression) in relation to objects and other humans can be automatically inferred, a predictive model could be built with enough data of what a person in any given situation might do. Combine this with robust speech recognition and understanding of the behaviour of common objects I think you would be closer to cracking the goal of language understanding (which I think defines true human level intelligence).
And i'd go so far as to say that this wouldn't just be good business, it would revolutionize human life.
But from what little I understand of the history of AI, people have been throwing resources at this category of problem for 40+ years and made only middling progress.
Not that that means that people (especially saavy entrepeneurs) shouldn't tackle the problem, I just think that by taking the constrained view (and shooting for the moon) you'd be missing a ton of low-hanging fruit in other areas.
With the emergence of commercially viable computing on demand we're tearing down a major barrier to AI, but I think that throwing the idea out and saying 'Hey! Why don't more of us get together and tackle this problem?' betrays a bit of an underestimation of just how tough this problem is.
That said, I would love for you to prove me wrong =) And as another poster mentioned, groups in the areas of surveillance and security (where the money is, at least initially) are making good headway on the problem.
The biggest problem with building predictive models, as another poster pointed out, is feeding them tagged training data. You could really feasibly load up a database with terabytes of data for most of the stimuli people react to these days (or will be able to in the near future), and even most of the basic combinations. It's a matter of tagging and classifying them in the first place that's the problem. Maybe with enough Indians sitting in front of mechanical turk... scrambles off to a calculator
But from what little I understand of the history of AI, people have been throwing resources at this category of problem for 40+ years and made only middling progress.
Not that that means that people (especially saavy entrepeneurs) shouldn't tackle the problem, I just think that by taking the constrained view (and shooting for the moon) you'd be missing a ton of low-hanging fruit in other areas.
With the emergence of commercially viable computing on demand we're tearing down a major barrier to AI, but I think that throwing the idea out and saying 'Hey! Why don't more of us get together and tackle this problem?' betrays a bit of an underestimation of just how tough this problem is.
That said, I would love for you to prove me wrong =) And as another poster mentioned, groups in the areas of surveillance and security (where the money is, at least initially) are making good headway on the problem.
The biggest problem with building predictive models, as another poster pointed out, is feeding them tagged training data. You could really feasibly load up a database with terabytes of data for most of the stimuli people react to these days (or will be able to in the near future), and even most of the basic combinations. It's a matter of tagging and classifying them in the first place that's the problem. Maybe with enough Indians sitting in front of mechanical turk... scrambles off to a calculator
That's exactly my initial reason for posting this. In the singulariy enthusiast community artificial intelligence is the most important thing we as humans can create (post scarcity, immortality etc).
What I was suggesting is that a "gold rush" type situation form around the building of classifiers, such as that occurred with website creation. There are millions trying to get an income onlne creating content for adsense and affilate programs (and for free with wikipedia and open source). Most will fail at generating reasonable income but in the process a large amount of wealth has been created. The same could be done for classifiers/artifical intelligence. It would require just collecting a thousand or two pieces of data (image/sudio/text) to create a classifier, an effort similar to creating a small website.
The problem is that even in the past there really wasn't a large number of people working on artificial intelligence, maybe just a few hundred or thousand, and computing power wasn't cheap enough to make it work. Computer scientists that do work on it are trying to create more unique algorithms that can be published, rather than creating applications with existing working algorithms (after all they''re paid to create new things, not simply apply what's available). In Larry Page's lecture at AAAS he mentioned that there are less than a hundred people working on artificial (general) intelligence today. (you can see it on youtube).
What I was suggesting is that a "gold rush" type situation form around the building of classifiers, such as that occurred with website creation. There are millions trying to get an income onlne creating content for adsense and affilate programs (and for free with wikipedia and open source). Most will fail at generating reasonable income but in the process a large amount of wealth has been created. The same could be done for classifiers/artifical intelligence. It would require just collecting a thousand or two pieces of data (image/sudio/text) to create a classifier, an effort similar to creating a small website.
The problem is that even in the past there really wasn't a large number of people working on artificial intelligence, maybe just a few hundred or thousand, and computing power wasn't cheap enough to make it work. Computer scientists that do work on it are trying to create more unique algorithms that can be published, rather than creating applications with existing working algorithms (after all they''re paid to create new things, not simply apply what's available). In Larry Page's lecture at AAAS he mentioned that there are less than a hundred people working on artificial (general) intelligence today. (you can see it on youtube).
What platform could support this vision? Also, how will you get all these disparate classifiers to integrate and scale without getting bogged down in complexity?
btw, on the computer vision example, I believe this is what Numenta is doing to an extent.
btw, on the computer vision example, I believe this is what Numenta is doing to an extent.
You can use open source libraries such as opencv/torch/torchvision and libraries in R for the machine learning. Haar recognizer is one such algorithm in opencv/torchvision that can locate faces in images.
As for the complexity, that's what the market would sort out. People would build on other's work, and everyone would get compensated for as much as their work is used.
Numenta is an example of the monolithic comapny doing all the work, and that's why you can't use or build on any of their technology yet (incidentally, they use bayesian type algorithms while i'm advocating easy to use standard algorithms that have many open source implementations - boosting, random forests, support vector machines, and maybe neural networks).
A startup opportunity would be for a company to host the learning algorithms, and just have people pay for and submit data from which classifiers (academic jargon for recongizers) can be built and returned. (The company would have to be trusted though as they could simply keep a copy of the created classifier and resell it).
As for the complexity, that's what the market would sort out. People would build on other's work, and everyone would get compensated for as much as their work is used.
Numenta is an example of the monolithic comapny doing all the work, and that's why you can't use or build on any of their technology yet (incidentally, they use bayesian type algorithms while i'm advocating easy to use standard algorithms that have many open source implementations - boosting, random forests, support vector machines, and maybe neural networks).
A startup opportunity would be for a company to host the learning algorithms, and just have people pay for and submit data from which classifiers (academic jargon for recongizers) can be built and returned. (The company would have to be trusted though as they could simply keep a copy of the created classifier and resell it).
Great idea. Start working on a prototype and make sure you apply in time for the winter YC session.
"... speech recognition is produced by only by two giant companies (Microsoft and IBM), ..."
and ATT ~ http://www.research.att.com/~ttsweb/tts/index.php#top
and ATT ~ http://www.research.att.com/~ttsweb/tts/index.php#top
Sure it could be done, but who is going to pay for it? At $1/1000 uses you won't even be able to cover your hosting fees.
At 30 frames per second you would start to earn dollars in just a few minutes of video. The price of 1 dollar per thousand would actually be too high for most applications - there might be thousands of object recognizers applied to each image.
A recongition task would take at most a second, to earn 100 dollars would require a little over a day of computing at that rate.
An alternative way would be for small companies/individuals to start producing small fragments of functionality and offering it in a per for use basis. (as in a cat recognizer in an image, or a recognizer for the word "hello" in an audiofile).
To create the above recognizer would require gathering data (such as hundreds/thousands of pictures of cats) and feeding it to a supervised learning algorithm like svm, boosting etc.
Thousands (perhaps millions) of people creating their own "classifiers" and charging for them (perhaps 1000 uses for 1 dollar) over the web might lead rapidly to seeing/thinking machines.
(A similar model can also be used to pose competition to entrenched monopoly software like mathematica/maya/autocad - thousands of people creating and charging for small bits of functionality)