Learning Generative Models of 3D Shapes and Environments

人工智能和计算机科学中最耐人寻味和反复出现的问题之一是，机器何时可以被认为拥有人类水平的智能。在其最著名的版本中，图灵测试通过机器进行自然语言对话的能力来判断机器的人性。2014年，一款名为尤金的聊天机器人被许多人认为通过了测试。那么，图灵测试是最好的选择吗？是不是太容易了？在一个鲜为人知的以艾达·洛夫莱斯夫人命名的“爱情测试”中，人们根据机器的创造力或独创性来评判机器。洛夫莱斯比图灵早了大约100年，经常被认为是世界上第一个计算机程序员。1843年，她评论说，在计算机能够产生一些原创的东西之前，不能认为它们拥有人类的智能，而这并不是它们被编程来做的。*机器真的能变得有创造力吗？这是最终的问题，长期的追求将是缩小机器和人类在创造力方面的差距。在短期内，作为一名计算机图形学研究员，我想首先解决一个更切实的问题：训练机器学习和执行3D形状和环境的生成性模型，在这些模型中，结果不必是创造性的。虽然经典图形学主要关注从显性的场景描述中合成逼真的图像，但在新的图形学时代，我们希望合成所有形式的视觉内容，其中输入可以是抽象的(例如文本)，也可以只由一组样本组成。*我目前的研究目标是推进数据驱动的视觉内容创作和几何深度学习，使“大3D数据”成为现实，实现我的四个V数据生成。也就是说，生成的数据应该是大容量的，具有跨类别的多样性，类别内的多样性，以及新颖性，最终是原创性，才能通过洛夫莱斯测试。在接下来的五年里，我将在开发和培训3D形状和环境的生成模型、丰富视觉内容和设计原型方面推进最先进的技术，以服务于AR/VR、机器人、教育、医疗、智能家居以及设计和制造中的应用。此外，我还将继续推动计算创造力的极限。

One of the most intriguing and reoccurring questions in AI and computer science is when a machine can be considered to possess human-level intelligence. In its best-known version, the Turing Test judges the humanness of a machine by its ability to make natural language conversations. In 2014, a chatbot named Eugene was considered by many to have passed the test. So is Turing Test the best choice? Is it too easy? In the lesser-known "Lovelace Test", named after Lady Ada Lovelace, machines are judged by their creativity or originality. Lovelace pre-dated Turing by about 100 years and is often credited as the world's first computer programmer. In 1843, she remarked that computers cannot be thought to possess human intelligence until they can generate something original, which they were not programmed to do.******Can a machine truly become creative? That is the ultimate question and the long-term pursuit would be to close the gap between machines and humans in creativity. In the shorter term, and as a computer graphics researcher, I want to tackle a more tangible problem first: to train machines to learn and execute generative models of 3D shapes and environments, where the outcomes do not have to be creative. While classical graphics mainly focuses on realistic image synthesis from explicit scene descriptions, in the new era of graphics, we wish to synthesize all forms of visual contents, where the inputs can be abstract (e.g., texts) or consist of only a set of exemplars.******My current research objective is to advance data-driven visual content creation and geometric deep learning, making "big 3D data" a reality and fulfilling my four V's for data generation. Namely, the generated data should be in large volume and with cross-category variety, intra-category variation, and novelty and ultimately, originality, to pass the Lovelace Test. In the next five years, I will advance the state of the art in developing and training generative models for 3D shapes and environments, enriching visual contents and design prototypes to serve applications in AR/VR, robotics, education, health, smart homes, and design and manufacturing. As well, I will keep pushing the boundary of computational creativity.