Continual Learning without Continual Training

Name: Continual Learning without Continual Training
Uploaded: 2026-10-17T23:20:24.513264+00:00
Duration: 309 s
Description: The Latent Concept PFN model replaces traditional continual training with continual inference. It learns a fixed latent space during meta-training based on shared concepts, which is then adapted to new domains and classes using only in-context evidence (a growing memory) without updating any model parameters.

The Latent Concept PFN model replaces traditional continual training with continual inference. It learns a fixed latent space during meta-training based on shared concepts, which is then adapted to new domains and classes using only in-context evidence (a growing memory) without updating any model parameters.

Open the full episode