Slashdot Mirror
Teaching Computers to See with Games
An anonymous reader writes "The Pittsburgh Post-Gazette has a story on Peekaboom, a two-player on-line game in which one player tries to get the other player to guess a word associated with an image, by revealing parts of the image one click at a time. From the article, "The process of revealing objects, or highlighting images within the larger context of the photo, is the sort of thing that researchers in computer vision must do to teach computers to see.""
Always pick a hint.
That adds 25 points to your score.
During the bonus round you get points for clicking the same spot as your teammate. Once numbers start appearing, keep clicking right there for maximum points.
Pass if the word looks difficult. Don't hesitate.
Pass if your partner passes, too. He probably has a good reason to.
Teaching, computers, games, yeah, fascinating... so, what's the deal with the moderation here? Why are there so few comments with scores over +3? My default is +5 and the whole front page right now shows *zero* comments at that level. Did they get real stingy with the mod points all of a sudden?
Dear Slashdot: next time you want to mess with the site, add a rich-text editor for comments.
Bypassing captchas?
There's an old story from the early neural net image recognition days that seems germane to this. A group of researcher were trying to train an artificial neural net to recognize military tanks that were partially hidden in forested scenes (this was the bad old Cold War days and spotting Soviet tanks in West German forests was the problem du jour). Pictures of natural forested scenes with and without tanks were used to train and test the system. It seemed to work very well on all the training and test data.
But when they tried the system on more images, it failed miserably. Further investigation revealed that, by accident, all of the "tank" pictures had been taken on cloudy days and all of the non-tank pictures had been taken on sunny days. The system had learned, and learned beautifully, how to recognize cloudy vs. sunny days.
The point is that the software was good enough to learn to recognize the difference between the two populations of images but that that difference wasn't the one intended by the people working on the system. In the same vein, I'm sure that Peekaboom will learn to distinguish between objects in images but whether it learns the actual object or just some incidental characteristic of that pocture of the object will require a very very good diversity of training pictures to avoid accidental, non-meaningful patterns in the image data.
I do wish them luck. Perhaps Peekaboom could create a distributed version of the training process in which others can both submit and help train on new objects/images. Letting others submit images and train the system would help diversify the training & testing data sets. Because some people will, no doubt, submit porn, I'm sure the system might become quite adept at recognizing the nether regions of the human body.
Two wrongs don't make a right, but three lefts do.
In some sense, Slashdot is the ultimate MMORPG. The comments system provides almost immediate feedback in the form of replies and moderation. Most posters can be categorized into some sort of stereotype.
Some posters like giving lots of information and opinion and getting lots of replies in return. They typically have a little background in what they are discussing, or have a very strong opinion on the subject. When they post and get modded up and have lots of replies, they have achieved a personal victory.
Other posters enjoy causing mayhem. They will typically post a comment taking a very odd stance towards a topic that many people feel strongly about, or they may post blatantly incorrect information on a topic that everyone is well-versed in. Their goal is not direct replies specifically, but rather that a heated debate follows from that first troll post. The best troll posters are those who can get both a slew of replies and start a flamewar. Moderation is a peripheral concern to these players, but they obviously prefer to be modded upwards rather than downwards.
Another player is the newb. This player simply doesn't get that the forum is a game populated by players much better than he. He posts replies in earnest to troll posts and karma whore posts, and may try to make on-topic jokes. This type of user is frequently seen making Star Wars references, posts about "42", and other stupid things that garner him neither karma nor respect from his peers. He is also frequently seen repeating Benjamin Franklin's worn out "those who would blah blah blah" catchphrase.
Finally there are the vermin of the forum. These will typically post off-topic comments about all sorts of strange fetish behavior. Whether it be the innocuous first posters or the ASCII art purveyors, these posters are not welcomed by most of the community. That they are able to stick around despite constant down-modding is a testament to their cockroach-like existence.
However without moderation, no one is interested in posting. The last few stories have only a handful of comments and most of them are posted by the vermin. The karma whores don't stick around because there is no payoff, the newbs are all gone because they follow the whores like flies on dog crap, and the trolls have no one to troll with the newbs and whores gone.
The healthy Slashdot ecosystem is significantly disturbed by this sudden lack of moderation.
You seem to logically following from a set of news story and extrapolating ideas. Unfortunately, the conclusions you are implying are, in actuality, a little backwards. Let me explain. Bayesian filtering wasn't developed to fight spam. It has been around, in theory, a long time before spam filtering, and spam filtering is just one application where it has reached prominence (especially in the Slashdot community). Saying that "a spam filter was used to learn to play chess" is really a misnomer. Bayesian filtering for chess has been tried 100s of times, long before spam was ever even an application of the filtering. It has been found to be insufficient, alone, in dealing with many classification and prediction problems, including chess and general vision algorithms.
There are fundamental flaws in the assumptions made in Bayesian filtering that are a bit technical, but I'll try to distill it here, for you. First, Bayes rule is essentially a way of turning the expression (A given B) into (B given A). The idea is if I give the filter a huge sample of preclassified data, it can build up probabilities of each event. In English, you can take 10,000 emails and find the quantity "the probability of this word occurring, given the message is spam". Bayes rule gives you a mathematical way of turning that sentence around, and finding "the probability this message is spam, given this word has occurred.".
In a single case, this is highly accurate. The problem cause when you begin to generalize this to entire email. Bayesian filters (in their simplest form) assume that each word is statistically independent of each other word. This is utterly absurd, of course, but in filtering email, it appears that this assumption doesn't cause too damage.
However, when you try to translate Bayesian filtering to computer vision, gigantic problems arise. First of all, what is your actual data? Color? Texture (Fourier or wavelet analysis)? Edge positions? Corner positions? How do you go about finding the more complicated ones? How do you deal with scale and rotation? None of these problems exist in email. The input is very simple, and already extremely easy for a computer to understand. An equivalent problem would be to classify a handwritten message as spam... now you have a whole new set of problems.
And, to further exacerbate the situation, if we take color as an example... the interdependencies between pixels is MUCH larger in an image, then the interdependencies of words in a message. Images are very coherent and predictable - ie, if I see a left eye, I can say with pretty large certainty there will be a right eye... or I'll know it's a profile, etc. There are huge and very macroscopic (difficult to quantify) dependencies that must be dealt with to form an effective classifier.
So, to answer your questions, Bayesian filtering is based upon a model of the data that is insufficiently modeling what is going on. Far too much information is neglected in a basic naïve bayes filter. Hope that helps.