Combining clustering and representation learning is one of the most promising approaches for unsupervised learning of deep neural networks. However, doing so naively leads to ill posed learning problems with degenerate solutions. In this paper, we propose a novel and principled learning formulation that addresses these issues. The method is obtained by maximizing the information between labels and input data indices. We show that this criterion extends standard crossentropy minimization to an optimal transport problem, which we solve efficiently for millions of input images and thousands of labels using a fast variant of the Sinkhorn-Knopp algorithm. The resulting method is able to self-label visual data so as to train highly competitive image representations without manual labels. Our method achieves state of the art representation learning performance for AlexNet and ResNet-50 on SVHN, CIFAR-10, CIFAR-100 and ImageNet.

Many systems and mechanisms in the human body are not fully understood, such as the principles of muscle control, the sensory nervous system that connects the brain and the body, learning in the brain, and the human walking motion. To address this knowledge deficit, we propose a human mimetic humanoid with an unprecedented degree of anatomical fidelity to the human musculoskeletal structure. The fundamental concept underlying our design is to consider the human mechanism, which contrasts with the conventional engineering approach used in the design of existing humanoids. We believe that the proposed human mimetic humanoid can be used to provide new opportunities in science, for instance, to quantitatively analyze the internal data of a human body in movement. We describe the principles and development of human mimetic humanoids, Kenshiro and Kengoro, and compare their anatomical fidelity with humans in terms of body proportions, skeletal structures, muscle arrangement, and joint performance. To demonstrate the potential of human mimetic humanoids, Kenshiro and Kengoro performed several typical human motions.

Clustering is a core building block for data analysis, aiming to extract otherwise hidden structures and relations from raw datasets, such as particular groups that can be effectively related, compared, and interpreted. A plethora of visual-interactive cluster analysis techniques has been proposed to date, however, arriving at useful clusterings often requires several rounds of user interactions to fine-tune the data preprocessing and algorithms. We present a multi-stage Visual Analytics (VA) approach for iterative cluster refinement together with an implementation (SOMFlow) that uses Self-Organizing Maps (SOM) to analyze time series data. It supports exploration by offering the analyst a visual platform to analyze intermediate results, adapt the underlying computations, iteratively partition the data, and to reflect previous analytical activities. The history of previous decisions is explicitly visualized within a flow graph, allowing to compare earlier cluster refinements and to explore relations. We further leverage quality and interestingness measures to guide the analyst in the discovery of useful patterns, relations, and data partitions. We conducted two pair analytics experiments together with a subject matter expert in speech intonation research to demonstrate that the approach is effective for interactive data analysis, supporting enhanced understanding of clustering results as well as the interactive process itself.

No abstract

No abstract

No abstract

This study investigates how pitch accent type and additive particles affect the activation of contrastive alternatives. In Experiment 1, German listeners heard declarative utterances (e.g., The swimmer wanted to put on flippers) and saw four printed words displayed on screen: one that was a contrastive alternative to the subject noun (e.g., diver), one that was non-contrastively related (e.g., pool), the object (e.g., flippers), and an unrelated distractor. Experiment 1 manipulated pitch accent type, comparing a broad focus control condition to two narrow focus conditions: with a contrastive or non-contrastive accent on the subject noun (nuclear L+H* vs. H+L*, respectively, followed by deaccentuation). In Experiment 2, the utterances in the narrow focus conditions were preceded by the unstressed additive particle auch (“also”), which may trigger alternatives itself. It associated with the accented subject. Results showed that, compared to the control condition, participants directed more fixations to the contrastive alternative when the subject was realized with a contrastive accent (nuclear L+H*) than when it was realized with non-contrastive H+L*, while additive particles had no effect. Hence, accent type is the primary trigger for signaling the presence of alternatives (i.e., contrast). Implications for theories of information structure and the processing of additive particles are discussed.

In order to realize more natural and various motions like humans, humanlike musculoskeletal tendon-driven humanoids have been studied. Especially, it is very challenging to design musculoskeletal body structure which consists of complicated bones, redundant powerful and flexible muscles, and large number of distributed sensors. In addition, it is very challenging to reveal humanlike intelligence to manage these complicated musculoskeletal body structure. This paper sums up life-sized musculoskeletal humanoids Kenta, Kotaro, Kenzoh and Kenshiro which we have developed so far, and describes key technologies to develop and control these robots