Trajectory replay improves lifelong learning in generative models
Trajectory Replay: Continuous-Time Mixture Transport for Lifelong Generative Models
Machine Learning
Summary
When computers learn to create images or data over time, they often forget what they learned first. The authors propose a new way called trajectory replay that helps these models remember old knowledge while learning new things by blending old and new information smoothly. This method avoids some technical problems in past approaches and works faster, needing fewer calculations. They tested it on image generation and biological data, showing it keeps memories better and makes useful predictions.
What this means in practice
- •For machine learning engineers: Improve image generation systems that need to learn new classes sequentially without forgetting previous ones by using trajectory replay.
- •For biomedical data teams: Extend generative models for biological data to new experimental conditions and batches while preserving prediction accuracy of treatment effects.
Authors
Alessandro Palma, Charles Jones, Tristan Cinquin, Fabian J. Theis, Frederik Wenkel
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.