Return of the normal distribution: Flexible deep continual learning with variational auto-encoders

Area of Science:

• Machine Learning
• Artificial Intelligence
• Deep Learning

Background:

• Continual learning from sequential data presents a significant challenge in machine learning.
• Existing deep learning solutions often oversimplify the problem for specific tasks, limiting generalizability.
• Variational Bayesian methods offer a more general perspective but have limitations.

Purpose of the Study:

• To investigate the utility of a generic variational auto-encoder (VAE) for continual learning.
• To adapt a two-stage training framework for continual learning contexts.
• To develop mechanisms for alleviating catastrophic forgetting in continual learning settings.

Main Methods:

• Utilized a generic two-stage variational auto-encoder (VAE) framework.
• Adapted the VAE for context-conditioned continual learning.
• Implemented generative rehearsal and data exemplar subset extraction to prevent forgetting.

Main Results:

• The proposed generic two-stage VAE demonstrated effectiveness in continual learning.
• The method surpassed task-tailored approaches in supervised classification.
• The VAE also excelled in unsupervised representation learning tasks.

Conclusions:

• A generic two-stage VAE provides a flexible and effective solution for continual learning.
• The proposed methods successfully address catastrophic forgetting without task-specific tailoring.
• This approach offers improved performance across diverse learning paradigms.