Gaussian process regression (GPR) or kernel ridge regression is a widely used and powerful tool for nonlinear prediction. Therefore, active learning (AL) for GPR, which actively collects data labels to achieve an accurate prediction with fewer data labels, is an important problem. However, existing AL methods do not theoretically guarantee prediction accuracy for target distribution. Furthermore, as discussed in the distributionally robust learning literature, specifying the target distribution is often difficult. Thus, this paper proposes two AL methods that effectively reduce the worst-case expected error for GPR, which is the worst-case expectation in target distribution candidates. We show an upper bound of the worst-case expected squared error, which suggests that the error will be arbitrarily small by a finite number of data labels under mild conditions. Finally, we demonstrate the effectiveness of the proposed methods through synthetic and real-world datasets.

While modern deep neural networks are performant perception modules, performance (accuracy) alone is insufficient, particularly for safety-critical robotic applications such as self-driving vehicles. Robot autonomy stacks also require these otherwise blackbox models to produce reliable and calibrated measures of confidence on their predictions. Existing approaches estimate uncertainty from these neural network perception stacks by modifying network architectures, inference procedure, or loss functions. However, in general, these methods lack calibration, meaning that the predictive uncertainties do not faithfully represent the true underlying uncertainties (process noise). Our key insight is that calibration is only achieved by imposing constraints across multiple examples, such as those in a mini-batch; as opposed to existing approaches which only impose constraints per-sample, often leading to overconfident (thus miscalibrated) uncertainty estimates. By enforcing the distribution of outputs of a neural network to resemble a target distribution by minimizing an$f$-divergence, we obtain significantly better-calibrated models compared to prior approaches. Our approach, f-Cal, outperforms existing uncertainty calibration approaches on robot perception tasks such as object detection and monocular depth estimation over multiple real-world benchmarks.

Object detection is a widely-used computer vision task, in which we identify a bounding box around each object in an image and classify the object into one of pre-defined classes. Single Shot MultiBox Detector (SSD) is a real-time object detector based on a single convolutional neural network. SSD is popular and known for high speed and accuracy, but its black-box nature is not ignorable when it is applied to critical systems. In this paper, we propose Contrastive Relevance Propagation (CRP), an extension of Layer-wise Relevance Propagation (LRP) tailored for SSD. CRP can consistently deal with SSD's heterogeneous output, i.e. confidences for object classes and location offsets, and create a heatmap that highlights a crucial part of the input, which is not available with a standard use of LRP. By experiments with the Pascal VOC 2012 dataset, we confirmed the quality of heatmaps created by CRP, and with such heatmaps, we conducted a simple analysis on prediction errors made by SSD.

In our previous study, we developed Pedagogical Agent with Gaze Interaction (PAGI), an anthropomorphic animated pedagogical agent that engages in gaze interaction with students. Using PAGI, we revealed that temporal contingency from virtual agents facilitate learning (temporal contingency effect), and proposed two hypotheses that may explain the result; 1) temporal contingency reduces extraneous cognitive load related to visual search, 2) temporal contingency prime social stance in learners which enhances learning. To assess more deeply into this matter, we tested two critical features of the agent, saliency and socialness. Two arrow shaped agents, of which differed in saliency, were employed. Apart from the appearance of the agents, the experimental design was identical to the previous study. University students learned words of a foreign language, with temporally contingent agent or recorded version of the agent, which played pre-recorded sessions from the contingent agents. From the result we gained evidence supporting the second hypothesis. Non-social agents did not trigger temporal contingency effect.