Moderator

Prof. Hao SUN, Tenured Associate Professor, GSAI

Session I

Chat and Language

09:15-09:45

Title：A Coarse Analysis of ChatGPT

Abstract：On December 1st last year, OpenAI launched ChatGPT, a new generation of conversational AI tool. ChatGPT has gained a lot of attention and buzz across the Internet for its amazing language understanding, generation and knowledge reasoning capabilities. The successful performance of ChatGPT has shown a possible path to solve Natural Language Processing, the core problem of cognitive intelligence. ChatGPT is considered a solid step towards the Artificial General Intelligence (AGI). ChatGPT will pose a huge challenge to applications such as search engines, and will even replace many people's jobs and disrupt many fields and industries. So, what scientific problem does ChatGPT solve, how does it solve that problem, and what other pressing problems will be solved in the future? We hope that the coarse analysis in this talk will partially answer these questions.

Prof. Wanxiang CHE, Professor, School of Computer Science and Technology, Harbin Institute of Technology

5mins

Title: Discussion on the Generalization of Large Language Model

Abstract：Large language models such as ChatGPT demonstrate powerful generality, as they can understand various user intents and complete multiple tasks. Their generality is mainly driven by the foundation model, instruction learning, and prompt learning. However, the generality of large models is still limited by the type, scale, and quality of training data, making it difficult to achieve a more comprehensive generality. In this report, We will first discuss the performance of large language models in terms of their generality, then introduce current enhancement techniques to expand the generality, and finally briefly share our attempts and explorations in this regard.

Prof. Jiajun ZHANG, Professor, National Laboratory of Pattern Recognition, Institute of Automation Chinese Academy of Sciences

5mins

Q&A

10:20-10:50

Title：Learning and Thinking on Conversational AI between Human and Computer

Abstract：Along with the emergence of ChatGPT, conversational AI has attracted unprecedented attention from both academia and industry. People now witness that the technology behind conversational AI is gradually accumulating, and resulting in fundamental changes. With deeper insights from researchers, it is possible to expect conversational AI in reality, not just Sci-Fi movies. This talk covers recent studies on topic modeling and data augmentation in dialogue systems, and also presents a brief introduction to ChatGPT as well as new challenges of conversational AI in the future.

Prof. Rui YAN, Tenured Associate Professor, GSAI

5mins

Q&A

10:55-11:25

Title：The Changing Focus of Large Language Model  Research

Abstract：In this talk, we will introduce changes in the research focus on natural language processing and large language models with technological development. We divide technological development into several stages, briefly introduce the research focus and main approaches of each stage, and focus on analyzing the important research directions in the current period of large language models.

Dr. Junlin ZHANG, Researcher, Sina Weibo

5mins

Q&A

11:30-12:10

Panel

12:30-13:30

Luncheon

Session II

Multimedia Content Generation

Moderator

Prof. Di Hu, Tenure-Track assistant professor, GSAI

13:30-14:00

Title: Breaking the Curse of Dimensionality in Generative Modeling: A Homotopic Approach

Abstract：Generative modeling for high-dimensional data, such as images and audio, is extremely challenging due to the curse of dimensionality. To overcome this difficulty, I introduce a homotopic approach inspired by numerical equation solving, which involves designing a homotopy of probability distributions that smoothly progresses from simple noise distribution to complex data distribution. I will present two families of approaches that rely on such homotopies: score-based diffusion models and consistency models. Both approaches use a differential equation to convert data to noise and learn to estimate the time reversal with deep neural networks. These models generate high-quality samples by following the homotopy from tractable distributions to the original data distribution, achieving state-of-the-art performance in many generative modeling benchmarks. They both allow for flexible neural network architectures and zero-shot image editing and exhibit complementary trade-offs.

Prof. Yang SONG, Tenure-Track assistant professor, Caltech

Member of Technical Staff, OpenAI

5mins

Q&A

14:05-14:35

Title：AIGC Technology and Application in XR

Abstract: In the past year, both XR and AIGC have developed rapidly. The development of XR is mainly due to the exploration of VR devices by many large companies, such as  Quest by Meta, Pico by Bytedance, etc. Based on the popularity of these devices, XR has many application scenarios in the fields of  live broadcasting, digital human, 3D content creation, etc. The explosion of AIGC technology will greatly improve the development of these topics. This talk mainly includes:

1 Introduction to application scenarios in XR

2 AIGC technology in these scenarios

Shilei WEN, Researcher, Bytedance

5mins

Q&A

14:40-15:10

Title:Beyond Temporal Activity Localization (TAL) in Long-Form Video

Abstract: Temporal activity localization (TAL) is an important task in video understanding especially for long-form videos. While much research has advanced the methodology used in this task, I will give an overview of the remaining challenges that hinder further performance. Moreover, I will elaborate on a relatively new task for the long-form video domain that goes beyond TAL, namely video language grounding.

Prof. Bernard Ghanem,Deputy Director, KAUST AI Initiative

5mins

Q&A

15:15-15:45

Title：AI for Music

Abstract：In this talk, I will introduce our two recent works about AIGC for music. 1) Given a video, a controllable music transformer model was proposed to generate music according to the user-specified music genre and instruments.  2) We further introduced music theory knowledge into the music generation model, collected the first video-music symbol data set for training, and proposed a retrieval-based metric to evaluate the matching degree between the generated music and video.

Prof. Si LIU, Professor, Beihang University

5mins

Q&A

15:50-16:20

Title: Imaginative AI on the EDGE: transforming species discovery, content creation, self-driving cars, emotional well-being, and more

Abstract:  Most existing AI learning methods can be categorized into supervised, semi-supervised, and unsupervised methods. These approaches rely on defining empirical risks or losses on the provided labeled and/or unlabeled data. Beyond extracting learning signals from labeled/unlabeled training data, we will reflect in this talk on a class of methods that can learn beyond the vocabulary that was trained on and can compose or create novel concepts. Specifically, we address the question of how these AI skills may assist species discovery, content creation, self-driving cars, emotional health, and more. We refer to this class of techniques as imagination AI methods, and we will dive into how we developed several approaches to build machine learning methods that can See, Create, Drive, and Feel. See: recognize unseen visual concepts by imaginative learning signals and how that may extend in a continual setting where seen and unseen classes change dynamically. Create: generate novel art and fashion by creativity losses. Drive: improve trajectory forecasting for autonomous driving by modeling hallucinative driving intents. Feel: generate emotional descriptions of visual art that are metaphoric and go beyond grounded descriptions. Feel: generate emotional descriptions of visual art that are metaphoric and go beyond grounded descriptions, and how to build these AI systems to be more inclusive of multiple cultures. . I will also conclude by pointing out future directions where imaginative AI may help develop better assistive technology for multicultural and more inclusive metaverse, emotional health, and drug discovery.

Prof. Mohamed Elhoseiny, Assistant Professor, Electrical and Mathematical Science and Engineering Division, KAUST

5mins

Q&A

16:25-16:55

Title：Fast, Controllable and Multi-modal Diffusion Models Toward Generative AGI

Abstract: Diffusion models have shown promise in various generation tasks including image generation, text-to-image generation, 3D scene generation and speech generation. Such models provide fundamental tools to build generative AGI, which is expected to understand and generate multi-modal data via a user-friendly interface like ChatGPT. Toward building such generative AI systems, this talk presents recent advances in fast and controllable inference algorithms, as well as unified backbone and probabilistic framework for multi-modal diffusion models.

Prof. Chongxuan LI, Tenure-Track assistant professor, GSAI

5mins

Q&A

17:00-17:30

Title: Modeling and Animating 2D Virtual Humans

Abstract: In recent years, 2D virtual humans have provided services in various scenarios including news reporting, teaching and entertainment. This talk gives an introduction to our recent advances in modeling and animating 2D virtual humans. For changing the appearance of virtual humans, I will introduce StyleSwap (ECCV 2022), which swaps and blends faces with high-fidelity. For face reenactment, I will introduce VPNQ (AAAI 2023) which drives virtual humans with videos. For audio-driven lip sync, I will introduce AV-CAT (Siggraph Asia 2022) and StyleSync (CVPR 2023), which seamlessly drive virtual humans to speak with audio.

Dr. Hang ZHOU,Researcher, Baidu

5mins

Q&A

17:35-18:15

Panel