Newest papers

Trajectory Replay: Continuous-Time Mixture Transport for Lifelong Generative Models

Alessandro De Palma, Charles Jones, Tristan Cinquin, Fabian Theis, Frederik Wenkel · 2026 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2610.26749 · cited by 0 works (OpenAlex) · open-access PDF

French

Adapter des modèles de génération de données en temps continu pré-entraînés à des données entrantes de manière séquentielle sans oublier les tâches antérieures est un défi central de l'apprentissage tout au long de la vie. Une stratégie de premier plan, le replay génératif, utilise un modèle de génération entraîné sur les tâches historiques pour augmenter sa propre mise à jour continue avec des échantillons synthétiques de l'expérience. Cependant, il nécessite des simulations de trajectoire complètes à chaque étape de formation, qui accumulent des erreurs de discrétion et peuvent dégrader la répartition historique répétée. Ici, nous proposons un replay de trajectoire, qui contourne la génération d'échantillons en formant un champ de vit

Machine translation (Meta M2M100); the English original follows.

Abstract

Adapting pre-trained continuous-time generative models to sequentially incoming data without forgetting earlier tasks is a central challenge in lifelong learning. A leading strategy, generative replay, uses a generative model trained on historical tasks to augment its own continual update with synthetic samples from experience. However, it requires full trajectory simulations at every training step, which accumulate discretization errors and can degrade the replayed historical distribution. Here, we propose trajectory replay, which bypasses sample generation by training an updated velocity field to match both a frozen historical velocity along its trajectories and the incoming task's velocity field, weighted by a stability-plasticity coefficient. Through the lens of continuous-time transport, we show the resulting optimum averages the two task velocities by their respective densities and generates the mixture of their probability paths. Combined with mean velocity distillation, trajectory replay accesses historical trajectories using only a few network evaluations per continual update. Empirically, it outperforms generative replay and distillation baselines on class-incremental image generation. In two biological case studies, trajectory replay extends generative perturbation models to new contexts and experimental batches across two data modalities while preserving treatment-effect prediction.

Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese

Topics: Trajectory, Computer science, Generative grammar, Generative model, Artificial intelligence, Forgetting

Sources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange