Task-Agnostic Continual Learning Using Online Variational Bayes With Fixed-Point Updates

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Catastrophic forgetting hinders neural network adaptability to evolving data distributions.
• Continual learning research often assumes known task boundaries, limiting real-world applicability.
• Task-agnostic continual learning, where boundaries are undefined, presents significant challenges.

Purpose of the Study:

• To approximate the intractable online Bayes update for neural network weights.
• To develop a continual learning algorithm capable of handling nonstationary data distributions.
• To address the limitations of existing methods in task-agnostic continual learning scenarios.

Main Methods:

• Derived novel fixed-point equations for online variational Bayes optimization.
• Utilized multivariate Gaussian parametric distributions for posterior approximation.
• Developed the FOO-VB algorithm, iterating the posterior through fixed-point equations.

Main Results:

• The FOO-VB algorithm effectively handles nonstationary data distributions.
• FOO-VB operates with a fixed neural network architecture.
• The method achieves superior performance compared to existing techniques in task-agnostic settings.
• FOO-VB does not require external memory or access to previous data.

Conclusions:

• FOO-VB offers a viable solution for continual learning in dynamic environments.
• The algorithm overcomes catastrophic forgetting without task boundary information.
• This research advances the field of task-agnostic continual learning.