|
Can someone give an example of a multiple instance learning scenario ? In what scenario will we get bags of instances with the class label following this `OR' property that defines MIL. That is, the label is positive even if only one instance is positive and negative only when all of them are negative. |
|
Any detection task that could be solved with a 'sliding window detector'. For example, finding a face in an image, a 2 second long clip of someone answering a phone within a 30 second video, or a few sentences on a given topic within a long document. In each of these tasks, you might have labels only on the image/video/document level, but want to detect more specific instances of the positive class. Also, the classical example is the Musk dataset, which motivated the first paper on multi-instance learning by Dietterich, et. al. In the Musk dataset, the task is to predict whether a molecule smells musky or not. Molecules naturally take on many different shapes, despite having the same atoms connected with the same topology. Each bag in the musk dataset consists of measurements taken on several shapes of one molecule. It is known which molecules smell musky but not which shapes of each molecule are responsible for the smell. If this is the case, then how can you apply the same model to each particular instance. Won't the model based on the whole document be different from the one based on one/two sentences ? Isn't the purpose of the learned model to classify each instance individually ?
(Dec 03 '10 at 17:31)
Aman
1
There is more than one approach to multi-instance learning. Not all methods are actually capable of learning to recognize individual instances. Generally the methods that do learn instance models say that they model of the document label is to take the OR of all the window labels inside the document. ie, if you are trying to make a detector of topic X, and any sentence in the document is about X, then you report that the whole document is about X.
(Dec 04 '10 at 21:45)
Ian Goodfellow
Most MIL work does seem to focus on building an instance classifier, even though the training data comes in the form of bags (on the assumption that it's somehow easier to get labels for collections than for individual items, and the goal is to learn despite the mismatch in granularity). A separate problem is to learn a bag classifier from training bags. I'm personally interested in any work that does operate at this level -- any pointers?
(Dec 04 '10 at 22:28)
Kiri Wagstaff
|
