Ant Group’s Research Institute Publishes Four NeurIPS 2022 Papers on Advanced Computer Vision and AI

Recently, the world‑leading AI conference NeurIPS 2022 announced its paper acceptance results, and the Ant Technology Research Institute, founded one year ago, had four papers accepted, focusing on frontier issues in next‑generation computer vision.

NeurIPS (Neural Information Processing Systems) is one of the most prestigious conferences in machine learning and computational neuroscience, covering topics such as machine learning, computer vision, natural language processing, and neuroscience; this year it received 10,411 submissions with an acceptance rate of 25.6%.

Amid the historic opportunity of digital transformation and technology‑driven innovation, Ant Group formally established the Ant Technology Research Institute in mid‑2021 to explore frontier science and technology. The institute comprises six labs: databases, graph computing, privacy computing, compiler, and visual intelligence. The four accepted papers come from the Visual Intelligence Lab, following two papers from the same lab that were accepted at ICML 2022.

The Visual Intelligence Lab conducts research on next‑generation computer‑vision fundamentals, including visual asset editing/creation (images, videos, 3D assets) for design assistance, human‑computer interaction, and low‑carbon social media; neural‑network interpretability and deep‑learning theory to understand model internals and improve trust; and large‑scale cross‑modal representation learning to bridge vision with other domains such as language, enhancing model generality and scalability.

The four selected papers are:

1. Rank Diminishing in Deep Neural Networks – Authors: Ruili Feng, Kecheng Zheng, Yukun Huang, Deli Zhao, Michael Jordan, Zheng‑Jun Zha. This work experimentally estimates the rank of each layer in common neural networks and reveals a rank‑diminishing induced independence loss, showing that class confidence can be linearly determined by a few other classes, advancing theoretical and empirical understanding of deep networks.

2. Improving 3D‑aware Image Synthesis with a Geometry‑aware Discriminator – Authors: Zifan Shi, Yinghao Xu, Yujun Shen, Deli Zhao, Qifeng Chen, Dit‑Yan Yeung. The paper proposes a 3D‑aware discriminator that extracts geometric information from images to supervise the generator, enabling fair adversarial training and significantly improving geometric quality and multi‑view continuity, with potential applications in content generation, AR/VR, etc.

3. Improving GANs with a Dynamic Discriminator – Authors: Ceyuan Yang*, Yujun Shen*, Yinghao Xu, Deli Zhao, Bo Dai, Bolei Zhou. It introduces two training strategies that dynamically adjust discriminator capacity without extra computation—expanding capacity when data is abundant and shrinking it when data is scarce—to achieve better discriminator solutions and consequently improve 2D/3D generators.

4. Uncertainty‑Aware Hierarchical Refinement for Incremental Implicitly‑Refined Classification – Authors: Jian Yang, Kai Zhu, Kecheng Zheng, Yang Cao. The paper addresses semantic granularity uncertainty in real‑world incremental learning by dynamically constructing semantic inheritance and conflict relations, guiding hierarchical refinement of incremental models for on‑demand learning based on user or environment needs.

This year Ant Group officially launched a four‑pillar ESG sustainable‑development strategy (digital inclusion, green low‑carbon, technological innovation, open ecosystem). The paper authors include many academic experts and young university researchers, such as Michael I. Jordan, a pioneer in machine learning and chair of Ant Group’s technology think‑tank, highlighting the institute’s open‑innovation research route and its internship program that integrates industry scenarios with academic talent.

In recent years, Ant Group has continuously invested in frontier technologies, focusing on privacy computing, blockchain, graph computing, distributed databases, and green computing as its five core “root” technologies. Public data shows a three‑year consecutive R&D investment growth rate exceeding 39%, with 63% of staff being technical talent, ranking sixth nationwide in private‑enterprise R&D spending in 2022.