用于交互式视频检索的教练式主动学习模型

Conventional active learning approaches for interactive video retrieval usually assume the query distribution is unknown, because it is difficult to estimate with only a limited number of labeled instances available. It is thus easy to put the system in a dilemma whether to explore the feature space in uncertain areas for a better understanding of the query distribution or to harvest in certain areas for more relevant instances. In this project, we propose a novel approach called coached active learning (CAL), which makes the query distribution predictable through training and thus avoids the risk of searching on a completely unknown space. The estimated distribution, which provides a more global view of the feature space, can be utilized to schedule not only the timing but also the step sizes of the exploration and the exploitation in a principled way. Furthermore, by integrating a domain adaption process into CAL, the model can demonstrate encouraging performance when moving to new testing domain which is different from where the model is trained. The constribtuions of this project include: 1) it makes the query distribution predictable; 2) it can balance the exploration and exploitation; 3) it is domain adaptive; 4) it is able to reduce the noise created in the labeling process. The research of this project will result in a more pricipled and computable model for interactive video search, as well as the algorithmatic and theoretical foundations for related applications.

主动学习方法中的非确定性策略是构建交互式视频检索系统的流行解决方案。但是，非确定性策略却因过于注重搜索未知空间而忽略已知空间中的相关实例，从而经常使用户失去对检索算法的信心而导致搜索夭折。该问题产生的原因是相关实例的分布无法提前预知，所以模型无法很好地分配"探索"和"利用"的时机和力度。本项目提出一个新颖的的主动学习模型，该模型在传统的主动学习上添加了一个"教练"过程来辅助学习。模型能动态估计相关实例在特征空间中的概率分布，从而像教练一样指导模型合理分配"探索"和"利用"的时机和力度。该模型的创新之处在于: 1)使相关实例分布具有可预测性,避免了传统的盲目式搜索;2)动态平衡探索和利用,使学习更有计划性;3)能根据测试集与训练集间的差异做自适应调整;4)能削弱人机交互过程中产生的噪声，具有较高的鲁棒性。该模型将为交互式视频检索建立规范化的可计算模型，并为相关应用提供算法和理论基础。

主动学习方法中的非确定性策略是构建交互式视频检索系统的流行解决方案。但是，非确定性策略却因过于注重搜索未知空间而忽略已知空间中的相关实例，从而经常使用户失去对检索算法的信心而导致搜索夭折。该问题产生的原因是相关实例的分布无法提前预知，所以模型无法很好地分配“探索”和“利用”的时机和力度。本项目提出一个新颖的的主动学习模型，该模型在传统的主动学习上添加了一个“教练”过程来辅助学习。模型能动态估计相关实例在特征空间中的概率分布，从而像教练一样指导模型合理分配“探索”和“利用”的时机和力度。该模型的创新之处在于: 1)使相关实例分布具有可预测性,避免了传统的盲目式搜索;2)动态平衡探索和利用,使学习更有计划性;3)能根据测试集与训练集间的差异做自适应调整;4)能削弱人机交互过程中产生的噪声，具有较高的鲁棒性。该模型将为交互式视频检索建立规范化的可计算模型，并为相关应用提供算法和理论基础。