JD Big Data R&D Department Presents Three Accepted Papers at AAAI-2018

JD Big Data R&D Department, in collaboration with institutions such as the Chinese Academy of Sciences, Northwestern University, and UC Berkeley, had three papers accepted at the 32nd AAAI conference (AAAI‑2018), marking the first time a single department secured three simultaneous acceptances at this top AI venue (acceptance rate 24.6%).

1. Cross‑domain Human Parsing via Adversarial Feature and Label Adaptation Authors: Si Liu, Yao Sun, Defa Zhu, Guanghui Ren, Jizhong Han, Jiashi Feng, Yu Chen This work proposes a cross‑domain adaptive human parsing model consisting of a feature compensation network, a feature adversarial network, and a structured label adversarial network. By using the LIP dataset as source domain and four target domains (surveillance video, movies, runway video, etc.), the method achieves strong performance, demonstrating how adversarial learning can enable semi‑supervised and transfer learning for large‑scale, weakly‑labeled image data.

2. Latent Discriminant Subspace Representations for Multi‑view Outlier Detection Authors: Kai Li, Sheng Li, Zhengming Ding, Weidong Zhang, Yun Fu The paper addresses outlier detection across multiple data sources, introducing three outlier types (category, attribute, and category‑attribute) and a novel unsupervised scoring metric. By learning a global low‑rank representation and formulating a constrained optimization problem, the method outperforms four baselines on five UCI datasets, enabling robust anomaly detection without any labeled data.

3. Orthogonal Weight Normalization: Solution to Optimization over Multiple Dependent Stiefel Manifolds in Deep Neural Networks Authors: Lei Huang, Xianglong Liu, Bo Lang, Adams Wei Yu, Yongliang Wang, Bo Li The study extends orthogonal matrix constraints from recurrent networks to general feed‑forward networks by modeling the learning problem as optimization over several dependent Stiefel manifolds. It introduces an orthogonal weight normalization technique that constructs orthogonal transformations for proxy parameters, ensuring stable gradient flow and improving training and generalization on CIFAR and ImageNet for Inception and ResNet architectures.

These contributions illustrate JD’s leading research in computer vision, machine learning, and deep learning, and underscore the department’s commitment to advancing AI technologies for applications ranging from intelligent monitoring to retail transformation.