CMU AI Learning Common Sense By Watching the Internet

An anonymous reader writes with this excerpt from the Washington Post "Researchers are trying to plant a digital seed for artificial intelligence by letting a massive computer system browse millions of pictures and decide for itself what they all mean. The system at Carnegie Mellon University is called NEIL, short for Never Ending Image Learning. In mid-July, it began searching the Internet for images 24/7 and, in tiny steps, is deciding for itself how those images relate to each other. The goal is to recreate what we call common sense — the ability to learn things without being specifically taught."

Re:Why not just go the obvious AI route-hard work

[GoodNewsJimDotCom]

If you're interested, I just opened a blog [botcraft.org] I think I'll pursue this to raise AI awareness.

Re:Deep Learning

[Anonymous Coward]

It has absolutely nothing to do with deep learning (DL).

DL is based on stacks or trees of classifiers where each top level classifier feeds lower levels. The idea here is that a classifier (say, a human face detector) can be built by smaller, much more specific (such as one for eyes, one for nose, one for hair, one for ears, etc), classifiers which are wrapped up by a larger classifier. This opposes the rather traditional approach of a single classifier for a whole bunch of data.

I believe the DL approach is inspired by random forests but I have yet to see Andrew Ng comment on that. Anyways, the cat research thingy was (semi)*SUPERVISED* learning. I.e.: here is a bunch of cat videos, there is a cat in them, learn what it is.

What TFA describes is *UNSUPERVISED* learning where the visual content and its meaning (written description) are inferred. I.e.: here is a bunch of random images followed by some not exactly descriptive text, learn the associations.

its not learning

[globaljustin]

this is just a program that analyzes text & images then returns sentences which humans can make sense from based on algorythm...*not saying its 'easy'* but its not a "thinking machine" or "learning common sense" in any way.

It is simply indexing the images & processing them according to the algorythm it was given.

TFA doesn't get into it much, but we can glean a bit from this:

Some of NEIL’s computer-generated associations are wrong, such as “rhino can be a kind of antelope,” while some are odd, such as “actor can be found in jail cell” or “news anchor can look similar to Barack Obama.”

that's the return...they define "common sense" as making associations between nouns and the images associated with the text on the origin page

"X can be a kind of Y"

analyze image

analyze text

identify nouns

associate nouns with image

idenfify all images that match noun

return: "X is related to Y"

"AI is a type of programmed computer response"...if you get my meaning ;)

Re:its not learning

[TapeCutter]

Coincidentally I came across the NEIL site last week, I think it has a long way to go before it can beat IBM's Watson on general knowledge (AKA "common sense"). Watson also gets it's raw information from the net, it categorises entities and discovers relationships between them. The difference is that Watson is not so much trained as it is corrected. Not unlike a human it can get a fundamental relationship or category wrong and that leads to all sorts of side-effects. In the Jeopardy stunt [youtube.com] they realised that humans had a slight advantage because they were informed when the other players made a right/wrong answer. When they gave Watson the same capability it was able to correctly identify the Jeopardy categories and then went on to convincingly beat the humans at their own game.

Computers are already better at "general knowledge" than humans despite the fact the "computer" needs 20 tons of air-conditioning to keep it running. The first time I saw the Jeopardy stunt it blew me away, my wife shrugged and said "So it's looking the answers up on the net. What's the big deal?". I can understand that from her since she has a Phd in marketing, what I don't understand is why most slashdotter's are similarly unimpressed? - I watched Armstrong land on the moon as a 10 year old boy but I think the history books will eventually give similar historical weight to Watson.

Re:Deep Learning

[TapeCutter]

Indeed. Personally I think IBM's "Watson" is the most impressive technological feat I have witnessed since I watched the moon landings 40-odd years ago, I fully realise few people share my amazement. The visual aspect means NEIL is tackling a far more difficult problem than deducing "common sense" from text alone. I wasn't impressed by the web site when I found it last week, but as a "proof of concept" it does the job admirably.

I may be wrong but I believe all three (Watson, NEIL, and the cat thingy) are based on the same general "learning algorithm" (neural networks, specifically RBM's). What they do is find patterns in data, both the entities (atomic and compound) and the relationships. The "training" comes in two types, feeding it specific facts to correct a "misconception" it has formed, labelling the entities and relationship it found so a human can make sense of it.

What the cat project did was train a neural net to recognise a generic cat by showing it pictures of cats and pictures of non cats. It could then categorise random pictures as either cat or not-cat, until fairly recently the problem has always been - How do I train the same AI to recognise (say) dogs without destroying it's existing ability to recognise cats.

Disclaimer: I knew the math of neural nets well enough 20yrs ago to have passed a CS exam. I never really understood it in the way a I understand (say) geometry but I know enough about AI and it's ever shifting goal posts to be very impressed by Watson's Jeopardy stunt. To convincingly beat humans at a game of general knowledge really is a stunning technological milestone that will be remembered long after 911 goes back to being just a phone number.