Model adapts to new tasks without retraining or forgetting old skills
Continual Learning without Continual Training
Machine Learning
Summary
Models that learn new things often forget what they learned before because they keep changing their settings. This paper’s authors propose a method where the model stops changing and instead learns what’s new just by remembering examples it’s seen. Their method uses a special way of thinking about concepts that helps it understand new ideas without losing old ones. This approach works for learning new types of data and new categories without confusion or forgetting.
What this means in practice
- •For machine learning engineers: Enable models to incorporate new classes or domains dynamically without retraining or tuning parameters, reducing deployment downtime.
- •For data platform teams: Maintain updated predictive models across evolving data streams by extending evidence sets instead of retraining, improving system robustness.
Authors
Nikita Narayanan, Ritham Majumdarr, Sonali Parbhoo
Abstract
Continual learning requires models to adapt to new domains and new classes while retaining prior knowledge. Many existing methods rely on continued optimization, using regularization, replay, or parameter expansion to prevent new updates from overwriting previously learned knowledge. Instead, we propose replacing continual training with continual inference: a PFN-based model that is meta-trained, and then frozen, adapting to new classes only by extending an in-context evidence set. Our model, Latent Concept PFN, performs in-context Bayesian inference over a latent concept space that captures semantic structure shared across domains and classes. As each new domain or class arrives, exemplars are added to the memory; adaptation reflects updated posterior beliefs over latent concepts rather than gradient updates. No parameters are changed, reducing forgetting. The same method handles both domain and class incremental continual learning without task identity. Concept annotations are only used during meta-training, acting as a soft anchor on the latent space rather than a fixed bottleneck. Unlike fixed-vocabulary concept methods, the model also handles noisy, ambiguous, or incomplete annotations by combining concept labels with raw input evidence to discover distinctions beyond the predefined concept set. Experiments on class and domain incremental learning datasets demonstrate competitive continual learning performance while learning interpretable latent concepts.