GESDA Anticipation Workshop - Future of AI Research

Afternoon

Martin Müller

Pascale Fung

As generative AI systems grow in complexity, increases in computational efficiency and better benchmarks for assessing progress and utility will be essential. A confluence of research in computer science, linguistics, ethics and other social sciences can help research move towards these goals, but the prime problem – intensive use of ever-diminishing training data resources – is likely to remain a pivotal challenge. What new paradigms might allow generative AI to further develop even in a data drought? What are the likely scenarios for generative AI in 5, 10 and 25 years from now?

Yejin Choi

Jun Zhu

AI systems that can construct detailed internal representations of their environment offer a number of advantages. Such “world modelling” enables systems to work in a more dynamic and effective fashion, updating the models, and associated parameters, as new information becomes available. It also makes for more robust reasoning and decision-making pathways, resulting in more transparent and more flexible planning capabilities. Achieving optimal world modelling will require progress in a number of areas, including neural-symbolic integration, continual learning and embodied AI. What are the breakthroughs we can expect to see in world modelling in the next 5, 10 and 25 years, and how will these affect the performance of AI systems at these junctures?

Yann LeCun

Jitendra Malik

AI has achieved some significant successes in science and mathematics, including helping scientists to further our understanding of protein folding and helping mathematicians to find new breakthroughs in knot theory. Scientists are currently applying AI to problems in superconductivity, genetics, turbulence, cosmology and drug discovery. What is the path forward here – will AI remain an assistant to scientists and mathematicians, or might it eventually make genuine discoveries unaided? How do we anchor our assessment of progress through robust benchmarks and ensuring that published achievements of AI for science are reliable and replicable? What will be the advances that occur at 5, 10 and 25 years in this domain?

Shirley Ho

Eric Xing

Maosong Sun

Recent breakthroughs in AI have been driven by algorithms that learn from data, in particular deep learning neural networks. Historically, training these models required datasets painstakingly labelled by humans. New model architectures like transformers now let models train on vast quantities of unlabelled data, greatly expanding the pool of training material. Progress in reinforcement learning, where models learn through experience rather than from curated datasets, is also expanding the scope of problems that AI can be applied to. However, many key ideas at the heart of the field are decades old and continued advances may require significant conceptual breakthroughs. The theory of deep learning also remains underdeveloped, which could both stall progress and limit our ability to understand models as they become more powerful. What significant theoretical breakthroughs in machine in learning can we expect in the next 5, 10 and 25 years? How will those breakthroughs affect the trajectory of AI research?

Kyunghyun Cho

Bernhard Schölkopf

Hang Li

Pascale Fung

GESDA Anticipation Workshop - Future of AI Research