Distributed and time-aware continual learning
Abstract
This research plan aims to produce human-like time-awareness in deep neural networks, which helps avoid catastrophic forgetting. We,propose a dynamically adaptive structural deep model with memory and an internal clock that stores the essence of old memories with,their corresponding time information and transfers the previous model parameters to a new task presented at a given time.We further,devise algorithms and protocols for distributed continual learning. We consider the setting where a number of continual learning age,nts can share information and learn from each other to improve their performance on newly arriving unseen tasks that other agents co,uld have seen. These protocols and algorithms are specially designed to address the issues brought by limited computation and communication resources.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Dec 10, 2021
- Source ID
- N000142212061
Entities
People
- Hoda Bidkhori
Organizations
- Office of Naval Research
- United States Navy
- University of Pittsburgh