On Sequential Bayesian Inference for Continual Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Continual learning aims to train models on sequential tasks without forgetting past knowledge.
• Sequential Bayesian inference offers a theoretical framework for continual learning by using previous posteriors as informative priors.
• Bayesian neural networks are explored for their potential in continual learning due to their probabilistic nature.

Purpose of the Study:

• To assess the effectiveness of sequential Bayesian inference in preventing catastrophic forgetting in Bayesian neural networks.
• To investigate the challenges of applying sequential Bayesian inference with Hamiltonian Monte Carlo in neural networks.
• To propose and evaluate a novel Bayesian continual learning method.

Main Methods:

• Sequential Bayesian inference was performed using Hamiltonian Monte Carlo.
• Posterior distributions were approximated using density estimators on Hamiltonian Monte Carlo samples.
• Analytical examples were used to study model misspecification and data imbalance effects.
• A new method, Prototypical Bayesian Continual Learning, was proposed and evaluated.

Main Results:

• Directly applying sequential Bayesian inference with Hamiltonian Monte Carlo failed to prevent catastrophic forgetting in Bayesian neural networks.
• Model misspecification and data imbalances were identified as significant challenges in continual learning.
• The proposed Prototypical Bayesian Continual Learning baseline achieved competitive results on class-incremental computer vision benchmarks.

Conclusions:

• Sequential Bayesian inference over neural network weights is insufficient for robust continual learning.
• Probabilistic models of the continual learning generative process are needed.
• Prototypical Bayesian Continual Learning presents a promising and effective approach for class-incremental continual learning.