Active Learning For Multi-Label And Multi-Modal Data

In traditional supervised learning problems,the system usually requires abundant labeled data to train reliable models.However,in many real-world learning tasks,the cost of labeling examples is rather high,especially for complex data with multiple labels or multiple modalities.And thus it becomes an important task to train an effective learning model with as few labeled examples as possible.Active learning,which actively selects the most valuable data to query their labels,is the most important approach to reduce labeling cost.In this paper,we focus on how to exploit active sampling techniques to ease the massive requirement of labeled data in different learning scenarios,especially for multi-label learning and cross modal similarity learning.In summary,our main contributions include:1.We propose a novel approach MADM for multi-label active learning by model guided distribution matching.On one hand,MADM exploits representativeness and diversity in both the feature and label space by matching the distribution between labeled and unlabeled data,and on the other hand,it tends to query predicted positive instances,which are expected to be more informative than negative ones.Experiments demonstrate that the proposed approach can reduce the labeling cost for multi-label data significantly.2.We propose a novel approach MIML-AL for multi-instance multi-label active learning by exploiting fine-grained supervised information.This algorithm designs a novel query strategy for multi-instance multi-label objects specifically and can acquire more precise information from the oracle without additional cost.Meanwhile,MIML-AL exploits diversity and uncertainty in both the input and output space to query the most valuable supervised information.Experiments show that our method can obtain the most improvement of learning model with the same labeling cost.3.We propose a novel approach COSLAQ for cross modal similarity learning with active queries.Based on the disagreement among different intra-modal and inter-modal similarities,COSLAQ explores the most valuable supervised information to improve the learning model.Furthermore,the closeness to decision boundary of similarity as uncertainty of metric learning is utilized to avoid querying outliers.And finally we validated its effectiveness empirically.