Understanding the Learning of Deep Networks: Expressivity, Optimization, and Generalization
|
Organizer(s): |
Name:
|
Affiliation:
|
Country:
|
Shijun Zhang
|
The Hong Kong Polytechnic University
|
Hong Kong
|
Feng-Lei Fan
|
The Chinese University of Hong Kong
|
Hong Kong
|
Juncai He
|
King Abdullah University of Science and Technology
|
Saudi Arabia
|
|
|
|
|
|
|
|
|
Introduction:
| | The rapid advancements in deep learning, demonstrated by technologies like ChatGPT and GPT-4, highlight the race towards artificial general intelligence. However, deploying these sophisticated AI systems raises severe concerns about risks, necessitating rigorous audits and oversight to ensure safety, robustness, and alignment with human values. Currently, managing an AI system is hindered by the lack of mathematical theory, making it challenging to intrinsically understand their learning functions, training processes, and generalizability.
To address these issues, the proposed special session will focus on the latest theoretical advancements in deep networks, covering expressivity, optimization, and generalization. The session will examine the capacity of networks to model complex functions, the implications of their non-convex optimization landscapes, and their ability to generalize and make accurate predictions about unseen data. This gathering will bring together leading researchers and practitioners to discuss recent progress, ongoing challenges, and future directions in the mathematical underpinnings of neural networks.
|
|
|
|
| |
