Continual Learning with Synthetic Boundary Experience Blending

TL;DR

This paper introduces Synthetic Boundary Data (SBD) generated via differential privacy to enhance experience replay in continual learning, leading to more stable decision boundaries and improved accuracy across multiple datasets.

Contribution

It proposes Experience Blending (EB), a novel framework that combines exemplars and synthetic boundary data through latent-space noise injection and joint training.

Findings

Achieves up to 13% accuracy improvement on Tiny ImageNet.

Enriches feature space near decision boundaries for better stability.

Demonstrates consistent gains across CIFAR-10, CIFAR-100, and Tiny ImageNet.

Abstract

Continual learning (CL) seeks to mitigate catastrophic forgetting when models are trained with sequential tasks. A common approach, experience replay (ER), stores past exemplars but only sparsely approximates the data distribution, yielding fragile and oversimplified decision boundaries. We address this limitation by introducing synthetic boundary data (SBD), generated via differential privacy: inspired noise into latent features to create boundary-adjacent representations that implicitly regularize decision boundaries. Building on this idea, we propose Experience Blending (EB), a framework that jointly trains on exemplars and SBD through a dual-model aggregation strategy. EB has two components: (1) latent-space noise injection to synthesize boundary data, and (2) end-to-end training that jointly leverages exemplars and SBD. Unlike standard experience replay, SBD enriches the feature space near decision boundaries, leading to more stable and robust continual learning. Extensive experiments on CIFAR-10, CIFAR-100, and Tiny ImageNet demonstrate consistent accuracy improvements of 10%, 6%, and 13%, respectively, over strong baselines.