|
|
A project of Visual Representation and Processing Group.
1School of Computing Informatics and Decision Systems Engineering, Arizona State University.
Multiple-instance learning (MIL) is a setting where labels are provided only for a collection of instances called bags. There are two types of instances: negative instances, which are found in either negative bags or positive bags, and positive instances, which are found only in positive bags. While a positive bag should contain at least one inherently positive instance, a negative bag must not contain any positive instances. In MIL, labels are not available at the instance level. It is interesting to note however that the label space is the same for both at the bag level and at the instance level. One may attempt to learn instance-level labels during the training stage, thus reducing the problem to an instance level supervised classification. Alternatively, one may also localize and prototype the positive instances in the feature space and rely on the proximity to these prototypes for subsequent classification.
MIL is an ideal set-up for many computer vision tasks. In particular, MIL can be an especially suitable model for medical image-based pathology classification and lesion detection localization, where an image is labeled pathological just because of one or a few lesions localized to small portions of the image. Medical images collected in a clinical setting may readily have an image-level label (either normal or various levels of pathology) while lacking the exact location of the lesion(s). The figure above illustrates such an example: color fundus images of eyes affected with different pathologies of diabetic retinopathy (DR). It is easy to notice that, although majority of the image looks normal, a small retinal landmark is enough to alter the label of the image from normal to pathological. In a MIL formulation for this problem, each image can be considered a bag and patches of images can be considered instances.
The following video explains the concept.
Each instance belonging to a particular cluster is independently drawn from a normal distribution that defines the said cluster. While positive bags can draw a subset of random cardinality of instances from negative distributions, negative bags cannot draw any data from positive distributions. Every positive bag must have at least one instance sampled from a positive distribution (marked in green ellipses P1 through P4, in the above figure). The centroids of these clusters would be the ideal positive instance prototypes that a MIL algorithm should identify. A prototype is an idealized point around which any point will share simliar characteristics.
In most MIL cases, one constructs a function as shown above that 'models' the 'positivity' of a MIL neighborhood. Any local maximas in this space are the positive instance prototype that we are looking for. Once a prototype is figured, its a straightforward manimization of training error to identify a radius around the prototype to construct a hyper-sphere that is a region for classifying a point as a truly and inherently positive instance.
If you have any questions/comments about any of these papers/code/datasets, please address them to Ragav Venkatesan