Improving Semantic Analysis on Point Clouds via Auxiliary Supervision of Local Geometric Priors

The Absence of a Weak-Tie Effect When Predicting Large-Weight Links in Complex Networks

Link prediction is a hot issue in information filtering. Link prediction algorithms, based on local similarity indices, are widely used in many fields due to their high efficiency and high prediction accuracy. However, most existing link prediction algorithms are available for unweighted networks, and there are relatively few studies for weighted networks. In the previous studies on weighted networks, some scholars pointed out that links with small weights play a more important role in link prediction and emphasized that weak-ties theory has a significant impact on prediction accuracy. On this basis, we studied the edges with different weights, and we discovered that, for edges with large weights, this weak-ties theory actually does not work; Instead, the weak-ties theory works in the prediction of edges with small weights. Our discovery has instructive implications for link predictions in weighted networks.

Geometry-Aware Generation of Adversarial Point Clouds

Machine learning models are vulnerable to adversarial examples. While most of the existing adversarial methods are on 2D image, a few recent ones extend the studies to 3D point clouds data. These methods generate point outliers, which are noticeable and easy to defend against using the simple technique of outlier removal. Motivated by the different mechanisms humans perceive by 2D images and 3D shapes, we propose the new design of geometry-aware objectives, whose solutions favor the desired surface properties of smoothness and fairness. To generate adversarial point clouds, we use a misclassification loss that supports continuous pursuit of malicious signals. Regularizing the attack loss with our proposed geometry-aware objectives results in our proposed method, Geometry-Aware Adversarial Attack (GeoA³). The results of GeoA³ tend to be more harmful, harder to defend against, and of the key adversarial characterization of being imperceptible. We also present a simple but effective algorithm termed GeoA₊³-IterNormPro towards surface-level adversarial attacks via generation of adversarial point clouds. We evaluate our methods on both synthetic and physical objects. For a qualitative evaluation, we conduct subjective studies by collecting human preferences from Amazon Mechanical Turk. Comparative results in comprehensive experiments confirm the advantages of our proposed methods. Our source codes are publicly available at https://github.com/Yuxin-Wen/GeoA3.