In the standard paradigm of Continual Learning (CL), a model learns a sequence of tasks. The primary enemy is catastrophic forgetting: as the model weights shift to minimize loss on Task B, they inevitably overwrite the features optimized for Task A. For years, the most effective defense has been "rehearsal"—maintaining a small buffer of old data (exemplars) to periodically remind the model of what it once knew.
However, in real-world deployments, this buffer is often a liability. Strict privacy regulations like GDPR, proprietary data-handling policies, and the limited storage of edge devices frequently prohibit the retention of historical training samples. This has birthed a specialized research direction: Exemplar-Free (or Data-Free) Continual Learning.
The goal is to enable a model to adapt to new domains, languages, or classes while maintaining performance on the past, using only the current task's data and the current model's weights.
The Geometry of Non-Interference
If we cannot use old data to "anchor" the model, we must ensure that new updates happen in a way that doesn't damage old knowledge. A recurring strategy is to find "safe" directions for gradient updates.
Researchers like Kaixun Huang and Enneng Yang propose using Orthogonal Projections. In Towards Rehearsal-Free Multilingual ASR and Continual Model Merging without Data, the intuition is that the model's parameter space contains "subspaces" important for specific tasks. By projecting new gradients into the null space of previous tasks, the model can learn new information without altering the weights critical for old ones. Enneng Yang extends this to model merging, treating the stability-plasticity trade-off as a multi-objective optimization problem where task vectors approximate the missing data subspaces.
Parameter Isolation and Adapters
Rather than moving the entire model, we can isolate the changes to a tiny fraction of the parameters. This is often achieved through Low-Rank Adaptation (LoRA) or Prompting.
- LoRA-based Isolation: James Seale Smith in ConStruct-VL uses a "Layered-LoRA" architecture to teach Vision-Language models structured concepts one by one. Similarly, Chaoyang Li introduces CaLoRA in Turning the Tables, which uses causal reasoning to identify which parameters are safe to update, even achieving "positive backward transfer"—where learning Task B actually improves performance on Task A.
- Prompting: Instead of changing weights, we can learn "instructions" for the model. Haoran Duan and colleagues in RefFiL use globally shared prompts in federated settings to maintain domain knowledge without clients ever sharing or storing private historical data.
Compensating for Feature Drift
When a model learns Task B, its internal "backbone" inevitably shifts. This causes feature drift: the representation of a Task A object (like a "cat") moves to a different part of the latent space, causing the Task A classifier to fail.
To solve this without old data, researchers use Drift Compensation. In Exemplar-Free Continual Representation Learning via Learnable Drift Compensation, Alex Gomez-Villa and Andrew D. Bagdanov propose learning a "translator" that predicts how features have moved.
Similarly, the CAMP framework by Grzegorz Rypeść and Bartłomiej Twardowski in Generalized Continual Category Discovery uses an auxiliary network to predict the drift of class centers. By combining this with Projected Distillation, they ensure that the new model's outputs remain consistent with the old model's logic, even as the underlying feature space evolves.
Generative Replay: Dreaming of the Past
If you aren't allowed to store real data, why not create your own? Generative Replay involves training a generator (or using the model's own generative capabilities) to produce synthetic examples of old tasks.
- Internal Replay: Andreas S. Tolias suggests a brain-inspired approach in Nature Communications, replaying hidden representations through feedback connections rather than raw pixels.
- VLM Self-Correction: In One VLM to Keep it Learning, the GaB framework leverages a Vision-Language Model’s ability to "dream up" questions about new images that relate to old tasks, creating a synthetic rehearsal buffer that satisfies privacy constraints.
Why It Works: The Power of Pre-training
A surprising finding in this field, highlighted by James Seale Smith in A Closer Look at Rehearsal-Free Continual Learning, is that the strength of the initial pre-trained model matters more than complex algorithms. When a model starts with a robust, general representation, simple L2 regularization can often outperform sophisticated prompting methods.
This suggests that exemplar-free learning is not just about "fixing" forgetting, but about preserving the semantic alignment established during pre-training. As Jinghuan Shang argues in Continual Learning with Global Alignment, anchoring representations to a shared, global semantic space provides a "north star" that keeps the model from drifting off course as it encounters new tasks.