Objective-Driven AI: Towards Machines that can Learn, Reason, and Plan

How could machines learn as efficiently as humans and animals? How could machines learn how the world works and acquire common sense? How could machines learn to reason and plan? Current AI architectures, such as Auto-Regressive Large Language Models fall short. I will propose a modular cognitive architecture that may constitute a path towards answering these questions. The centerpiece of the architecture is a predictive world model that allows the system to predict the consequences of its actions and to plan a sequence of actions that optimize a set of objectives. The objectives include guardrails that guarantee the system's controllability and safety. The world model employs a Hierarchical Joint Embedding Predictive Architecture (H-JEPA) trained with self-supervised learning. The JEPA learns abstract representations of the percepts that are simultaneously maximally informative and maximally predictable. The corresponding working paper is available here.

Frontier of AI

The research and development of AI have reached a watershed moment in the history of the technology. Traditionally, theoretical and leading research results came from university labs while the industry produced more applied research translating them into actual products. However, the new generation of frontier AI models are coming out of industry labs due to the requirement of huge amounts of training data and computing resources that are unavailable to most university research labs. In addition, many of the frontier models are not open to the public, and therefore not shareable with the academic community. In light of this, what is the future model of industry-academia collaboration? What should today’s AI PhD students work on in terms of their thesis topic when the SOTA seems to change every few months? What can we look forward to in the next 3 years in terms of AI research trends? What are the important AI problems to solve? We will discuss these and other topics in this fireside chat with Turing Award Laureate Professor Yann LeCun, Member of the US National Academy of Engineering Professor Harry Shum, moderated by Professor Pascale Fung, AAAI, ACL and IEEE Fellow, Director of the Centre for AI Research (CAiRE) at HKUST.

Addressing Large Language Models that Lie: Case Studies in Summarization

The advent of large language models promises a new level of performance in generation of text of all kinds, enabling generation of text that is far more fluent, coherent and relevant than was previously possible. However, they also introduce a major new problem: they wholly hallucinate facts out of thin air. When summarizing an input document, they may incorrectly intermingle facts from the input, they may introduce facts that were not mentioned at all, and worse yet, they may even make up things that are not true in the real world. In this talk, I will discuss our work in characterizing the kinds of errors that can occur and methods that we have developed to help mitigate hallucination in language modeling approaches to text summarization for a variety of genres.

Pengcheng Mind and AI Large Model Cooperation

As a revolutionary chat pre-training model, ChatGPT has already placed a huge impact on global economic. It is the strong foundation of computing power that enables large models to continuously improve in the process of understanding massive data, resulting in breakthrough innovations. Based on the Pengcheng Cloud Brain E-level intelligent computing platform, Pengcheng Laboratory is training Pengcheng Mind foundation model, which is the first fully controllable, safe, open-source pre-training foundation model in China, with a parameter level of 200 billion, performance benchmarking with ChatGPT, and output in line with Chinese core values. In the future, Pengcheng Laboratory will open up Pengcheng Mind cooperation and work with external partners to continuously build a large model open-source consortium for domestic large model ecosystem.

Holistic Artificial Intelligence (HAI)

Today, we have access to hundreds of large generative models and millions of smaller, task-specific models. Despite this, we currently lack a mechanism that allows for flexible dispatching and composition of multiple models to tackle the myriad of intelligent tasks we encounter in real-world scenarios.

In this talk, Dr. Feng will unveil a novel framework, namely “Holistic Artificial Intelligence” (HAI), designed to alleviate the difficulty of AI service development by orchestrating available AI models. With HAI, a user or client can articulate his/her intelligent service requirement in a multitude of natural ways, such as through natural language, speech, illustrative images, action sequences, or a combination of these methods. The central control unit of HAI powered by foundation models maps this requirement to an execution plan consisting of a specific model or a cascaded process of multiple models. Additionally, it aligns these models with the most suitable computing and network facilities for efficient deployment. This talk will discuss the four major technical challenges that underpin the HAI framework, related work in the community, and her team’s explorations into this exciting field.

Exploring the New Frontiers of AI – ByteDance's Pursuits

The emergence of ChatGPT has ushered in a new era of artificial intelligence (AI), offering unprecedented possibilities. The immense potential of the large AI models, especially those developed recently, is poised to revolutionize both our personal lives and professional activities. Along with these advancements also come novel challenges, issues, and opportunities. In this talk, Li will shed light on the ongoing research initiatives at ByteDance Research. Their mission is to spearhead the development of cutting-edge technologies in this new AI era. Currently, ByteDance Research is dedicating its efforts to several areas, including robotics, AI for science, responsible AI, and autonomous agents. During the talk, he will introduce the ultimate objectives, key technologies, and noteworthy outcomes of these research endeavors.

Experience and Challenges in Building Large Language Models at Alibaba

In this talk, the speaker will share his experience and lessons learned during the development of Tongyi Qianwen, also known as Qwen, a state-of-the-art large language model at Alibaba Cloud. He will first outline the key steps taken to construct a high-performing model with the ability to generate creative text, comprehend intricate instructions, tackle mathematical problems, and more. Subsequently, he will describe a variety of systems challenges in building large language models and present their innovative design in the areas of distributed storage, high performance networking, resource scheduling, and execution frameworks. Such techniques significantly enhance the efficiency of handling complex AI workloads in a distributed environment.