Compact Goal Representation Learning via Information Bottleneck in Goal-Conditioned Reinforcement Learning.

We propose an Information bottleneck (IB) for Goal representation learning (InfoGoal), a self-supervised method for generalizable goal-conditioned reinforcement learning (RL). Goal-conditioned RL learns a policy from reward signals to predict actions for reaching desired goals. However, the policy would overfit the task-irrelevant information contained in the goal and may be falsely or ineffectively generalized to reach other goals. A goal representation containing sufficient task-relevant information and minimum task-irrelevant information is guaranteed to reduce generalization errors. However, in goal-conditioned RL, it is difficult to balance the tradeoff between task-relevant information and task-irrelevant information because of the sparse and delayed learning signals, i.e., reward signals, and the inevitable task-relevant information sacrifice caused by information compression. Our InfoGoal learns a minimum and sufficient goal representation with dense and immediate self-supervised learning signals. Meanwhile, InfoGoal adaptively adjusts the weight of information minimization to achieve maximum information compression with a reasonable sacrifice of task-relevant information. Consequently, InfoGoal enables policy to generate a targeted trajectory toward states where the desired goal can be found with high probability and broadly explores those states. We conduct experiments on both simulated and real-world tasks, and our method significantly outperforms baseline methods in terms of policy optimality and the success rate of reaching unseen test goals. Video demos are available at infogoal.github.io.

Region Anomaly Detection via Spatial and Semantic Attributed Graph in Human Monitoring.

This paper proposes a graph-based deep framework for detecting anomalous image regions in human monitoring. The most relevant previous methods, which adopt deep models to obtain salient regions with captions, focus on discovering anomalous single regions and anomalous region pairs. However, they cannot detect an anomaly involving more than two regions and have deficiencies in capturing interactions among humans and objects scattered in multiple regions. For instance, the region of a man making a phone call is normal when it is located close to a kitchen sink and a soap bottle, as they are in a resting area, but abnormal when close to a bookshelf and a notebook PC, as they are in a working area. To overcome this limitation, we propose a spatial and semantic attributed graph and develop a Spatial and Semantic Graph Auto-Encoder (SSGAE). Specifically, the proposed graph models the "context" of a region in an image by considering other regions with spatial relations, e.g., a man sitting on a chair is adjacent to a white desk, as well as other region captions with high semantic similarities, e.g., "a man in a kitchen" is semantically similar to "a white chair in the kitchen". In this way, a region and its context are represented by a node and its neighbors, respectively, in the spatial and semantic attributed graph. Subsequently, SSGAE is devised to reconstruct the proposed graph to detect abnormal nodes. Extensive experimental results indicate that the AUC scores of SSGAE improve from 0.79 to 0.83, 0.83 to 0.87, and 0.91 to 0.93 compared with the best baselines on three real-world datasets.

Hierarchical Gaussian Descriptors with Application to Person Re-Identification.

Describing the color and textural information of a person image is one of the most crucial aspects of person re-identification (re-id). Although a covariance descriptor has been successfully applied to person re-id, it loses the local structure of a region and mean information of pixel features, both of which tend to be the major discriminative information for person re-id. In this paper, we present novel meta-descriptors based on a hierarchical Gaussian distribution of pixel features, in which both mean and covariance information are included in patch and region level descriptions. More specifically, the region is modeled as a set of multiple Gaussian distributions, each of which represents the appearance of a local patch. The characteristics of the set of Gaussian distributions are again described by another Gaussian distribution. Because the space of Gaussian distribution is not a linear space, we embed the parameters of the distribution into a point of Symmetric Positive Definite (SPD) matrix manifold in both steps. We show, for the first time, that normalizing the scale of the SPD matrix enhances the hierarchical feature representation on this manifold. Additionally, we develop feature norm normalization methods with the ability to alleviate the biased trends that exist on the SPD matrix descriptors. The experimental results conducted on five public datasets indicate the effectiveness of the proposed descriptors and the two types of normalizations.