Here’s how image recognition works in a nutshell. It starts with identifying points of interest in an image — the points, lines, and patterns that provide sharp contrasts or really stick out from a bland, featureless background. It’s similar in some ways to how the human eye picks out edges and points by keying off the places where there’s sharp contrast.
Then it looks at how these points are related to each other — the geometry of the whole set of points. You could picture it as looking like a constellation of stars, even though really it’s a more sophisticated mathematical model of these points of interest and how they relate.
Now it compares that model to all the other models in a huge database. Those other models come from images it has already analyzed from around the web. It looks for a matching model, but it doesn’t have to be a perfect match. In fact, it’s important that it be a bit flexible, so it doesn’t matter if it’s turned around, or shrunken, or twisted a bit. The Taj Mahal still has the basic geometry of the Taj Mahal even if you photograph it from a little bit of a different angle or photograph it lower in the frame. When Google recognizes that it matches that model best, it guesses it’s probably the Taj Mahal.
They do use SIFT (or at least a variant thereof) for finding and describing interest points, but by itself, there is no geometric matching in SIFT. There are various competing approaches on how do it, although in many cases, you can get very good results even without it. (It's very slow to do geometric matching so people often skip that step, or only apply it to the best matches.)
Landmark detection is a recent "hot topic" in computer vision, and given a large enough dataset, it essentially works now for the most part.
Sorry if this adverts the discussion, but where are we with AI in weather prediction? It seems to me that we're mostly still using an older model with scientists analyzing models to find patterns, but machine learning and weather forecasting seems like a big field for AI improvements.
I don't think we need AI to do forecasts, because we can already do a direct simulation of all the parameters at once. Sure, it takes a supercomputer, but there's enough money in to to still be profitable. So why take shortcuts?
I was under the impression that weather models are at least partially chaotic. Is this not so?
In such systems, even if we have a perfect model we cannot know the initial conditions with sufficient accuracy to make reliable long term predictions. This is not just a limit of our current technology, but a property of chaotic systems. If weather is a chaotic system, then we will never make serious progress beyond short term forecasts.
... you can get 'attractors', basins in the dynamics. If either several different models, or slightly tweaked starts with the same model give you the same or similar-enough answers, then you know you've stumbled into a basin by luck. Small errors aren't going to change much. If your outcomes are different though, then you know that the weather from that starting state is very sensitive and unpredictable.
A change in the initial conditions might make a change in the result, but it isn't guaranteed to. At least not any noticeable change in the next few days, weeks, even months. So if you find that, for most of the possible current conditions, the end result is mostly the same, you can have a good probability of being right.
Someone should sponsor a contest to develop a weather forecasting algorithm like Netflix did for their movie preferences algorithm. Verifying the answer should not be hard! :)
“Only about 15% of what we do is traditional forecast meteorology,” says CEO Bill Kirk
I once had a brief conversation with a meteorology student about this. The idea of learning a model from example data seemed foreign to him. It seemed to me that there could be things to improve by applying ML algorithms.
The challenge Google faces on how to improve the way users retrieve information from the web is actually limited by its own search interface. Google's simple search box interface has successfully trained most users to abstract complex queries into few keywords, and as result, much context are lost as we are searching. In some way, Google Instant is a way to mitigate this problem by previewing results per keywords, but the search experience is still limited to a keyword index, regardless the interface Google wraps around with.
What's interesting moving forward how AI can help search is about how NOT to discard the constraints of human language. In ML, constraints are good, and for search, it can help the indexer to more intelligently quantify the n-grams. End of day, human being still converse in sentences, not keywords.
Information retrieval and information extraction is evolving, thanks to statistical machine learning theories.
Successful applications of statistical pattern recognition models improve user experience (speech recognition, face tracking, sentiment analysis, and other countless examples). Other parts of Artificial Intelligence have just benign effect.
Here’s how image recognition works in a nutshell. It starts with identifying points of interest in an image — the points, lines, and patterns that provide sharp contrasts or really stick out from a bland, featureless background. It’s similar in some ways to how the human eye picks out edges and points by keying off the places where there’s sharp contrast.
Then it looks at how these points are related to each other — the geometry of the whole set of points. You could picture it as looking like a constellation of stars, even though really it’s a more sophisticated mathematical model of these points of interest and how they relate.
Now it compares that model to all the other models in a huge database. Those other models come from images it has already analyzed from around the web. It looks for a matching model, but it doesn’t have to be a perfect match. In fact, it’s important that it be a bit flexible, so it doesn’t matter if it’s turned around, or shrunken, or twisted a bit. The Taj Mahal still has the basic geometry of the Taj Mahal even if you photograph it from a little bit of a different angle or photograph it lower in the frame. When Google recognizes that it matches that model best, it guesses it’s probably the Taj Mahal.