[Paper Review] Nonparametric Bayesian Structure Adaptation for Continual Learning.

Abhishek Kumar, Sunabha Chatterjee+1 more|arXiv (Cornell University)|Dec 8, 2019

Domain Adaptation and Few-Shot Learning4 citations

TL;DR

This paper proposes a nonparametric Bayesian method that dynamically grows neural network architecture during continual learning by learning task-specific sparse weight subsets. It enables inter-task transfer through overlapping weights and achieves state-of-the-art performance on benchmark datasets while automatically adapting network structure without prior architectural constraints.

ABSTRACT

Continual Learning is a learning paradigm where machine learning models are trained with sequential or streaming tasks. Two notable directions among the recent advances in continual learning with neural networks are (i) variational Bayes based regularization by learning priors from previous tasks, and, (ii) learning the structure of deep networks to adapt to new tasks. So far, these two approaches have been orthogonal. We present a principled non-parametric Bayesian approach for learning the structure of feed-forward neural networks, addressing the shortcomings of both these approaches. In our model, the number of nodes in each hidden layer can automatically grow with the introduction of each new task, and inter-task transfer occurs through the overlapping of different sparse subsets of weights learned by different tasks. On benchmark datasets, our model performs comparably or better than the state-of-the-art approaches, while also being able to adaptively infer the evolving network structure in the continual learning setting.

Motivation & Objective

• Address the limitations of existing continual learning methods that either fix network architecture or use parametric priors that do not adapt to task complexity.
• Enable automatic growth of hidden layer nodes as new tasks are introduced, avoiding manual architecture design.
• Facilitate effective inter-task knowledge transfer by learning overlapping sparse subsets of weights across tasks.
• Develop a principled Bayesian framework that jointly learns model structure and parameters in a continual learning setting.

Proposed method

• Employ a nonparametric Bayesian prior over network weights, allowing the number of active neurons in each layer to grow adaptively with new tasks.
• Use a sparse Bayesian learning framework where each task learns a distinct, sparse subset of weights, with overlapping weights enabling transfer.
• Leverage variational inference to approximate the posterior over network weights and structure, enabling scalable training.
• Introduce a hierarchical prior structure that encourages sparsity and promotes shared representations across tasks.
• Dynamically infer the number of active neurons per layer based on task complexity and data distribution.
• Ensure continual learning by updating the posterior distribution incrementally as new tasks arrive, without catastrophic forgetting.

Experimental results

Research questions

• RQ1 Can a nonparametric Bayesian model automatically adapt the network architecture to evolving task demands in continual learning?
• RQ2 How effectively can overlapping sparse weight subsets enable knowledge transfer between tasks without explicit architectural constraints?
• RQ3 Does the proposed method outperform existing parametric and nonparametric approaches in terms of accuracy and structural adaptability?
• RQ4 To what extent does the model prevent catastrophic forgetting while growing in complexity over time?

Key findings

• The model achieves performance comparable to or better than state-of-the-art continual learning methods on standard benchmark datasets.
• The network structure dynamically expands with new tasks, with the number of active neurons in each layer growing organically as needed.
• Inter-task transfer is effectively enabled through overlapping sparse weight subsets, improving generalization across tasks.
• The method successfully prevents catastrophic forgetting by maintaining and updating a posterior distribution over weights and structure.
• The approach demonstrates scalability and robustness in sequential learning scenarios without requiring architectural redesign.