Modular Neural Networks

Meeting URL: https://nus-sg.zoom.us/j/87981040210?pwd=akdUVVE0c01VSC9ZWXdDTW1IaHhwQT09
Meeting ID: 879 8104 0210
Password: 754915


Abstract:
The term Modular Network was known before its application to Neural Networks and was defined multiple times in different areas of research, including Applied Mathematics, Social Sciences, Biology, and others.
Li and Shuurmans defined modularity as a function quantifying the quality of a network division into communities.
So Modular Networks (MN) -- are the networks containing highly connected regions (communities), which are sparsely connected to the rest of the network.

In this work we study the evolvement of different types of Modular Networks in the field of deep learning. The focus of our research is in the advantages of such architectures applied to different domains and tasks.
We will start by introducing the advantages of modular architectures in addressing high-dimensional output space problems, as well as proposing a novel uncertainty-based strategy for the structure inference.
Next, we will study the application of MNs to the problem of incremental learning and functionality disentanglement.
Finally, we will discuss advantages of modular architectures in sparse-reward problems of reinforcement learning.