Yitao Xu (徐一涛)

Welcome to my homepage! I am currently a 2nd year PhD student at École polytechnique fédérale de Lausanne (EPFL) in the EDIC program. Before that, I was a master's student at KTH Royal Institute of Technology, majoring in Machine Learning. In 2022 Fall and 2023 Spring, I was an exchange master's student in Computer Science at EPFL.

I got my B.Eng degree in Computer Science and Technology at Beihang University (BUAA). I worked as a research student at State Key Laboratory of Software Development Environment (NLSDE), advised by Prof.Xianglong Liu, and I also finished my bachelor's thesis project focusing on rectifying texture bias exposed in deep neural network there. After graduating from BUAA, I became a Machine Learning Intern at DiDi AI Labs, where I was led by Dr. Zhengping Che and Prof.Jian Tang. Beginning in 12.2020, I spent a wonderful winter at Tsinghua Laboratory of Brain and Intelligence (THBI) under the supervision of Prof.Jia Liu, studying intuitive physics engine and stability inference with convolutional neural networks.

Email  /  CV  /  Linkedin  /  Google Scholar  /  Github

I'm interested in Computer Vision and Machine Learning. Particularly, I focus on:

• Neural cellular automata and self-organizing systems
• Interpretability of adversarial attacks and adversarial robustness
• Different cognition biases between humans and deep-learning based computer vision models

We propose AdaNCA, inserting NCA into ViTs as plug-and-play adaptors for enhancing their robustness. We develop a unified perspective of NCA and ViT in terms of token interaction learning. Our work takes the first step to scale NCA up to solve large-scale computer vision problems, and proves that NCA can be practically utilized to improve the robustness of current vision transformer models.

MeshNCA is a comprehensive framework for 3D texture synthesis. It is compatible with image or text targets, and enjoys various test-time properties such as texture density control, grafting, and mesh generalization. Play with our demo!

We examine the behaviors of NCA at the spatio-temporal limit and find it struggling to generate correct patterns. We propose a simple yet effective solution by initializing the seeds with noises. The increased spatio-temporal continuity enables several test-time controls, including anisotropic scaling, multi-scale pattern formation, and continuous speed adjustment. Play with our demo!

We investigate the relationship between the NCA architecture and the emergent dynamics of the trained models.

We propose Dynamic Neural Cellular Automata (DyNCA), a framework for real-time and controllable dynamic texture synthesis. Play with our demo!

We believe that adversarial noises are amplified and propagated through the critical attacking route, identified by the proposed algorithm.

We take the first step to study adversarial attacks for embodied agents. We generate spatiotemporal perturbations to form 3D adversarial examples, exploiting the interaction history in both the temporal and spatial dimensions.

We explain adversarial robustness for deep models from a new perspective of neuron sensitivity which is measured by neuron behavior variation intensity against benign and adversarial examples.

Design