OUseful.Info, the blog…

Trying to find useful things to do with emerging technologies in open education

People Powered Supervised Training Algorithms: Google Does it Again?

If you ever do a course on artificial intelligence or machine learning, you are likely to presented with the idea of supervised learning. In a supervised learning algorithm, training examples are presented to a system being trained to perform some sort of classification task, along with information about the class that the training example falls into. The learning algorithm uses this desired output to reward the classifier if it produces the correct output classification for a given input, or punish it otherwise. (The strength of the reward/punishment may also depend on how close the actual output is to the desired output for a given input.)

Machine learning algorithms like these have been used for a long time in the context of machine learning, but what do you do you training set does not contain examples of the the correct answers? How then can you supervise the training of the system?

In the GWAP – Games With a Purpose [PDF] – approach pioneered by Luis von Ahn, people are used to provide the training signal (von Ahn also calls this approach Human Computation…)

Whenever you use the Google search engine, you potentially contribute to the training of the Google search algorithm. Each click on a result can be used to train the search algorithm that result (or that advert) is potentially a good match for the current search term.

Recently, Google bought ReCaptcha, publishers of a popular system for checking that there’s a human on the end of a browser by presenting them with two hard to read words and getting the user to type them in. The clever thing about ReCaptcha is that the machine itself doesn’t necessarily know what one of the words says – it’s using the human to teach it. But you knew that already right? After all, Charlie (aged 14) does…

So when I saw Google’s new toy – Google Building Maker – announced today, watched through the promo video:

and even had a play, I saw it not as “a very smart way for Google to enhance the 3D experience in Google Earth” (ReadWriteWeb) or as “crowdsourcing building making to their users” (TechCrunch), I saw it as a way of helping train Google image processing edge detection algorithms so that they can automate the creation of 3D models from multiple satellite images…

If you try out Google Building Maker, you’ll soon see how it could even start to use the information gathered from your own models in the first few images presented to you to start suggesting where to place the lines in the later images…

Written by Tony Hirst

October 13, 2009 at 7:01 pm

Posted in Anything you want

Tagged with

3 Responses

Subscribe to comments with RSS.

  1. […] Hirst sees Building Maker as an example of human “supervised learning” for improving artificial intelligence. He has several other interesting examples. […]

    Create 3D buildings online

    October 14, 2009 at 1:01 pm

  2. […] One thing to note is that it may take some time for someone to tweet what someone has said. If we had a transcript caption file (i.e. a timecoded transcript of the presentation) we might be able to work out the “mean time to tweet” for a particular event/twitterer, in which case we could backdate timestamps to guess the actual point in the video that a person was tweeting about. (I looked at using auto-genearated transcript files from Youtube to trial this, but at the current time, they’re rubbish. That said, voice search on my phone was rubbish a year ago, but by Christmas it was working pretty well, so the Goog’s algorithms learn quickly, especially where error signals are available. So bear in mind that if you do post videos to Youtube, and you can upload a caption file, as well as helping viewers, you’ll also be helping train Google’s auto-transcription service (because it’ll be able to compare the result of auto-transcription with your captions file…. If you’re the Goog, there are machine learning/supervised learning cribs everywhere!)) […]

  3. […] a master of casting applications so that they can benefit from supervised training (see for example People Powered Supervised Training Algorithms: Google Does it Again?) so with their weight behind it, the Prediction API could be an early indicator of a path that will […]

Comments are closed.


Get every new post delivered to your Inbox.

Join 1,349 other followers

%d bloggers like this: