Joint Clustering and Classification for Multiple Instance Learning.

Fig 1: Illustrates the underlying idea in Embedding Space based MIL approaches. A beach scene, segmented into regions, is represented as a bag of instances, where each instance represents the corresponding region. A set of concepts is then used to calculate a similarity between each instance and the concept, referred to as concept-wise instance similarity (CIS). The max MIL assumption is used to embed each bag into the concept space using the CIS. Classification can then be performed in the embedding space using standard classifiers (best viewed in color). This work proposes to learn both the concept space and the classifier in a joint fashion.

Project Synopsis: The Multiple Instance Learning (MIL) has been extensively used to solve weakly supervised visual classification problems, where visual data is represented as a bag of instances. Previous MIL algorithms mostly extended standard supervised learning algorithms to the MIL settings by defining the classification score of a bag as the maximum score of each of its instances. Although these algorithms (referred to as Instance Space (IS) based methods) are popularly used, they do not account for the possibility that the instances may have multiple intermediate concepts. On the other hand, Embedding-space (ES) based MIL approaches are able to tackle this issue by defining a set of concepts, and then embedding each bag into a concept space, followed by training a standard classifier in the embedding space.
Contribution: In previous ES based approaches, the concepts wered iscovered separately from the classifier, and thus were not optimized for the final classification task. This work proposes a novel algorithm to estimate concepts and classifier parameters by jointly optimizing a classification loss. This approach discovers a set of discriminative concepts, which yield superior classification performance. The proposed algorithm is referred to as Joint Clustering Classification for MIL (JC2MIL) because the discovered concepts induce clusters of data instances. Moreover, we show that the proposed algorithm achieves state-of-the-art results on several MIL datasets, by discovering fewer number of concepts compared to previous ES-based methods.

Method/Idea: Our target is to jointly learn (1) a set of concepts that are used to embed each bag into a concept space, and (2) a classifier that combines the embedding to produce a classification score. We achieve this by posing the problem as joint minimization of the classification loss with respect to both the set of concepts and the classifier parameters. The model makes two assumptions (1) the probability of a concept lying in a bag is maximum over the probability of each of its instances (similar to IS based methods), and (2) the similarity between kth concept and an instance isdefined using the rbf kernel (similar to [2, 3, 5] . The classification lossincludes the mean negative log-likelihood and a regularization term, and is written as:

where C and w are the concept and classifier parameters, and p_i is the classifier score for each bag. The above optimization is solved using coordinate descent.

Related publication: Sikka, K., Giri, R., Bartlett, M. (2015). Joint Clustering and Classification for Multiple Instance Learning. British Machine Vision Conference (BMVC15). [PDF] [short_abstract]

Results: Shown results on four MIL datasets. Pl. refer to publication for more details.

Code: The code can be downloaded at link. The code has been tested and distributed before. Thanks.