I am working on an experimental research project that addresses catastrophic forgetting and incremental learning efficiency. Current AI paradigms often rely on scaling massive models with billions of parameters to integrate new knowledge. While effective initially, this approach can be inefficient and computationally expensive for incremental learning.
In this project, the model is designed as a system of interconnected models (modules), where each module learns about a specific domain, maintaining its knowledge independently but coordinated by a Global Bayesian Node.
Problems it aims to solve:
- Catastrophic forgetting
- Scalability of knowledge modules
- Efficiency in incremental learning
I welcome suggestions, feedback, and contributions, as well as researchers working on similar problems or who have explored related challenges to share their ideas and experiences.
For more conceptual details and architecture, you can visit this website of the project.