We investigated human understanding of different network visualizations in a large-scale online experiment. Three types of network visualizations were examined: node-link and two different sorting variants of matrix representations on a representative social network of either 20 or 50 nodes. Understanding of the network was quantified using task time and accuracy metrics on questions that were derived from an established task taxonomy. The sample size in our experiment was more than an order of magnitude larger (N = 600) than in previous research, leading to high statistical power and thus more precise estimation of detailed effects. Specifically, high statistical power allowed us to consider modern interaction capabilities as part of the evaluated visualizations, and to evaluate overall learning rates as well as ambient (implicit) learning. Findings indicate that participant understanding was best for the node-link visualization, with higher accuracy and faster task times than the two matrix visualizations. Analysis of participant learning indicated a large initial difference in task time between the node-link and matrix visualizations, with matrix performance steadily approaching that of the node-link visualization over the course of the experiment. This research is reproducible as the web-based module and results have been made available at: https://osf.io/qct84/.

Beyond achieving high performance across many vision tasks, multimodal models are expected to be robust to single-source faults due to the availability of redundant information between modalities. In this paper, we investigate the robustness of multimodal neural networks against worst-case (i.e., adversarial) perturbations on a single modality. We first show that standard multimodal fusion models are vulnerable to single-source adversaries: an attack on any single modality can overcome the correct information from multiple unperturbed modalities and cause the model to fail. This surprising vulnerability holds across diverse multimodal tasks and necessitates a solution. Motivated by this finding, we propose an adversarially robust fusion strategy that trains the model to compare information coming from all the input sources, detect inconsistencies in the perturbed modality compared to the other modalities, and only allow information from the unperturbed modalities to pass through. Our approach significantly improves on state-of-the-art methods in singlesource robustness, achieving gains of 7.8-25.2% on action recognition, 19.7-48.2% on object detection, and 1.6-6.7% on sentiment analysis, without degrading performance on unperturbed (i.e., clean) data.

No abstract

No abstract

tw. 1slee~eate.sinica.edu.hv ABSTRCT Tone recognition for fluent Mandarin speech has always been a very difficult problem, because the pitch contours vary seriously with the context conditions and the complicated tone behavior is difficult to analyze. In this paper, a new set of four inter-syllabic features are identified to characterize quantitatively such pitch contour variation with respect to the context conditions. In addition, a robust pitch extraction method is proposed by integrating the Adaptive Gabor Representation (AGR) and Instantaneous Frequency Amplitude Spechum (IFAS). Experimental results indicate that accurate pitch values can be extracted under various noisy conditions, and the tone recognition accuracy can be improved significantly.