Reproduce “Multi-scale Feature Alignment for Continual Learning of Unlabeled Domains” on OpenResearch

About this paper

Kevin Thandiackal, Luigi Piccinelli, Pushpak Pati, Orcun Goksel

A framework for continual unsupervised domain adaptation enables deep learning models to adapt to a sequence of unlabeled domains without storing prior data, mitigating catastrophic forgetting by continuously updating a generative model to synthesize features from all past domains. The method achieves state-of-the-art performance for tissue-type classification on histopathological datasets, improving mean F1 scores by 6.9 percentage points over the second-best method, and effectively aligns feature spaces while preserving class discriminability.

Key insights

• Leveraging a single, dual-purpose discriminator for both generative adversarial network training and adversarial domain alignment efficiently promotes realistic feature generation and domain-invariant feature extraction.
• Aggregating features from multiple depths of the classifier (Multi-scale Feature Aggregation) allows the model to balance the need for general representations in generation with task-specific domain invariance in adaptation.
• Continuously updating a conditional generator to synthesize features from all past domains, combined with image and logit distillation, is crucial for comprehensively mitigating catastrophic forgetting across all previously encountered domains without storing real data.

Summary by alphaXiv.