Continual Learning without Continual Training
The paper proposes continual learning by freezing the model and updating only its in-context memory.
Latent Concept PFN adapts to new classes or domains by adding exemplars, not by changing weights. The method performs Bayesian inference over a latent concept space meant to preserve prior knowledge and reduce forgetting. Concept annotations are used during meta-training as a soft guide, while raw inputs still help handle noisy or incomplete labels. The authors report competitive results on class- and domain-incremental benchmarks with interpretable latent concepts. ArXiv · AI/CL/LG's note
Latent Concept PFN adapts to new classes or domains by adding exemplars, not by changing weights. The method performs Bayesian inference over a latent concept space meant to preserve prior knowledge and reduce forgetting. Concept annotations are used during meta-training as a soft guide, while raw inputs still help handle noisy or incomplete labels. The authors report competitive results on class- and domain-incremental benchmarks with interpretable latent concepts. ArXiv · AI/CL/LG's note
score 5