Rank-loss support instance machines for MIML instance annotation

Abstract

Multi-instance multi-label learning (MIML) is a framework for supervised classification where the objects to be classified are bags of instances associated with multiple labels. For example, an image can be represented as a bag of segments and associated with a list of objects it contains. Prior work on MIML has focused on predicting label sets for previously unseen bags. We instead consider the problem of predicting instance labels while learning from data labeled only at the bag level. We propose Rank-Loss Support Instance Machines, which optimize a regularized rank-loss objective and can be instantiated with different aggregation models connecting instance-level predictions with bag-level predictions. The aggregation models that we consider are equivalent to defining a "support instance" for each bag, which allows efficient optimization of the rank-loss objective using primal sub-gradient descent. Experiments on artificial and real-world datasets show that the proposed methods achieve higher accuracy than other loss functions used in prior work, e.g., Hamming loss, and recent work in ambiguous label classification.

Witness identification in multiple instance learning using random subspaces

“…Birds: The birds data set was introduced in [24]. In this data set, each bag corresponds to a 10 seconds recording of bird songs from one or more species.…”

Score thresholding for accurate instance classification in multiple instance learning

“…Birds [8]: In this data set, each bag corresponds to a 10 seconds recording of bird songs from one or more species. The recording is temporally segmented, and each part corresponds to a particular bird, or to background noises.…”

“…In the past, it has been used for molecule activity prediction [2], image classification [3], computer aided diagnosis [4], visual object tacking [5] and document classification [6]. MIL research traditionally focuses on bag classification, however, more recently, several authors considered problems in which instance must be classified individually [4], [7], [8].…”

Disambiguation-Free Partial Label Learning

“…The BirdSong data set 5 contains 4998 singing syllables recorded from 13 bird species, where each syllable is described by 38 features with candidate labels being bird species jointly singing during a 10-second period [4,17]. The MSRCv2 data set 6 contains 1758 image segmentations from 23 classes of objects, where each image segmentation is described by 48 his- togram and gradient features with candidate labels being objects appearing within the same image [17,22].…”