[Paper Review] Overcoming Catastrophic Forgetting by Generative Regularization

Patrick H. Chen, Wei Wei+2 more|arXiv (Cornell University)|Dec 3, 2019

Domain Adaptation and Few-Shot Learning40 references4 citations

TL;DR

This paper proposes Bayesian Generative Regularization (BGR), a memory-free continual learning method that combines variational Bayesian inference with energy-based generative modeling to mitigate catastrophic forgetting. By enforcing diverse feature learning through contrastive loss in an energy-based model, BGR improves posterior approximation and achieves state-of-the-art performance, outperforming prior methods by 15% on Fashion-MNIST and 10% on CUB.

ABSTRACT

In this paper, we propose a new method to overcome catastrophic forgetting by adding generative regularization to Bayesian inference framework. Bayesian method provides a general framework for continual learning. We could further construct a generative regularization term for all given classification models by leveraging energy-based models and Langevin-dynamic sampling to enrich the features learned in each task. By combining discriminative and generative loss together, we empirically show that the proposed method outperforms state-of-the-art methods on a variety of tasks, avoiding catastrophic forgetting in continual learning. In particular, the proposed method outperforms baseline methods over 15% on the Fashion-MNIST dataset and 10% on the CUB dataset

Motivation & Objective

• Address catastrophic forgetting in continual learning under memoryless, fixed-model settings where previous data cannot be re-accessed.
• Identify the limitation of mean-field variational inference in Bayesian continual learning, particularly the failure to maintain diverse, stable features across tasks.
• Improve posterior approximation in Bayesian neural networks by enforcing sufficient, holistic feature representations through generative regularization.
• Enable robust continual learning without storing past data or expanding model size, aligning with privacy regulations like GDPR.
• Demonstrate that generative modeling enhances feature diversity and stability, reducing forgetting in incremental task learning.

Proposed method

• Introduce a Bayesian generative regularization (BGR) framework that combines discriminative learning with implicit generative modeling via energy-based models (EBMs).
• Use Langevin dynamics sampling to approximate the generative loss, enabling the model to learn diverse features across all parts of input data.
• Incorporate a contrastive loss in the EBM to encourage the model to reconstruct input data accurately, promoting holistic feature learning.
• Integrate the generative regularization term into the variational inference objective, ensuring stable and diverse posterior approximations.
• Maintain a fixed model architecture throughout continual learning, avoiding parameter expansion or data replay.
• Leverage the generative capability to validate that features are not concentrated on discriminative parts, but spread across the entire input structure.

Experimental results

Research questions

• RQ1 Why does mean-field variational inference in Bayesian continual learning fail to prevent catastrophic forgetting despite its streaming update property?
• RQ2 Can generative modeling improve the quality of posterior approximation in Bayesian neural networks for continual learning?
• RQ3 Does enforcing holistic feature learning through generative regularization reduce forgetting in incremental task learning?
• RQ4 How does the proposed method compare to state-of-the-art memory-free and memory-based continual learning approaches in terms of accuracy and stability?
• RQ5 Can the generative regularization be effectively combined with existing memory-based methods to further improve performance?

Key findings

• The proposed BGR method reduces catastrophic forgetting by promoting diverse feature learning through generative regularization, as evidenced by uniformly distributed salient features across input strokes.
• On the Fashion-MNIST dataset, BGR improves accuracy by 15% over state-of-the-art methods, demonstrating strong performance in continual learning.
• On the CUB dataset, BGR achieves a 10% improvement in accuracy compared to baseline methods, highlighting its effectiveness on fine-grained classification tasks.
• Qualitative analysis using integrated gradients shows that BGR models distribute salient features evenly across digit strokes, whereas standard SGD models focus on isolated discriminative parts.
• The accuracy drop after learning a new task is significantly reduced—from 54.2% to 95.0%—when using BGR, indicating better retention of prior knowledge.
• BGR can be effectively combined with memory-based methods such as iTAML, improving performance from 97.8% to 98.4% on split-MNIST, suggesting broad compatibility and enhancement potential.