From the ACM: Learning More About Active Learning
The April edition of Communications of the ACM has an interesting article on recent advances in active learning by Graeme Stemp-Morlock.
In passive learning (a more traditional approach), you build a large training set of classified data by (often) manually assigning labels. This data is used as the basis of your analysis.
In the real world, we find that generating these large sets of labeled data is often expensive and time consuming. With active learning, you identify the most ambiguous data to label, resulting in a much higher payoff for each label defined (and fewer headaches for your labelers).
The article goes on to mention that active learning is being used in practice with excellent results (for example in music identification, text classification, and even bioinformative), but that the theory lags. This is another example of a gap between the world of the practitioner and the academic work behind it.
2 comments
Can you give us an example of this process?
How funny, I just made the tweet about Active Learning & the ACM publishes something on it. Guess I’m on the ball.
Need to subscribe so I can give it a read.
Leave a Comment