In the standard framing of continual learning, an agent encounters a sequence of distinct tasks—say, first learning to recognize birds, then cars, then faces. Most algorithms rely on a "task boundary" signal: a notification that Task A has ended and Task B has begun. This signal is the trigger for the model to consolidate its knowledge, freeze certain weights, or update a replay buffer.
However, as Chen Zeno, Elad Hoffer, and colleagues argue in Task-Agnostic Continual Learning Using Online Variational Bayes With Fixed-Point Updates, this assumption is a luxury that real-world systems rarely enjoy. In natural environments, data distributions shift gradually, overlap, or change without warning. This has led to a research direction focused on task-agnostic or task-free continual learning, where models must maintain stability and plasticity autonomously.
The Problem of the Hidden Shift
When task boundaries are removed, the "stability-plasticity dilemma" becomes significantly harder. Without a signal to trigger protection mechanisms, a model might inadvertently overwrite critical features learned from previous data distributions.
Researchers like Jisoo Mok and Seunghak Yu point out in Large-scale Lifelong Learning of In-context Instructions and How to Tackle It that real-world data streams are often "blurry." If a model cannot detect when it is entering a new domain, it cannot know when to expand its capacity or when to protect its current state. This leads to "catastrophic forgetting" not just of old tasks, but of the foundational features required to distinguish between old and new classes—a phenomenon Yiduo Guo and Bing Liu term Cross-Task Class Discrimination (CTCD) in Dealing with Cross-Task Class Discrimination in Online Continual Learning.
Bayesian Uncertainty as a Natural Lock
One powerful intuition for solving this is to treat the model's weights not as fixed points, but as probability distributions. In Variational Bayes for Federated Continual Learning, Dezhong Yao and colleagues use Bayesian Neural Networks (BNNs) to handle evolving distributions.
The logic is elegant: as the model learns, it becomes more "certain" about the weights that are crucial for past data (the variance of those weight distributions decreases). When new data arrives, the Bayesian framework naturally penalizes changes to these "certain" weights. Zeno et al. take this further with FOO-VB, using fixed-point updates to "lock" weights based on the curvature of the loss landscape. By looking at the geometry of the error surface, the model can sense which directions of change would be most damaging to its existing knowledge, effectively creating its own internal task boundaries.
Geometric and Functional Stability
If we cannot rely on a signal to tell us when to protect knowledge, we can instead design the model's internal space to be more resilient to change.
- Feature Projection: In Continually Learning Self-Supervised Representations with Projected Functional Regularization, Alex Gomez-Villa and Bartłomiej Twardowski propose Projected Functional Regularization (PFR). Instead of storing old data, they use a "temporal projection" to map new features back into the old feature space, ensuring that the model's "understanding" of the world remains consistent even as it learns new representations.
- Expanding Targets: Wanhao Yu and Yang Li address the problem of expanding capacity in Rethinking Continual Learning with Progressive Neural Collapse. Their ProNC framework dynamically grows the geometric "target" for features (a Simplex Equiangular Tight Frame) as new classes appear. This allows the model to integrate new information without collapsing the structure it built for previous classes.
Gradient Alignment in Decentralized Streams
The challenge of task-agnostic learning is amplified in decentralized settings, like Federated Learning. Jy-yong Sohn and Suman Banerjee introduce Fed-A-GEM in Buffer-based Gradient Projection for Continual Federated Learning.
In this scenario, different devices might experience different distribution shifts at different times. Fed-A-GEM works by projecting local updates against a "global buffer gradient." If a proposed update to the model would move in a direction that contradicts the collective knowledge stored in the buffer, the update is re-projected. This ensures that the global model only learns things that are "compatible" with what it has already mastered, regardless of whether a task boundary was ever declared.
Anchoring with Foundation Models
Finally, the rise of large-scale foundation models has provided a new "anchor" for task-agnostic learning. In POP: Prompt Of Prompts for Continual Learning, Dashan Gao and Nuno Vasconcelos use frozen models like CLIP. Instead of updating the entire network, they learn a sequence of prompts. By maintaining a "global" prompt alongside task-specific ones, the system can navigate a continuous stream of data by simply learning how to "query" its own frozen knowledge base, bypassing the need for explicit task-switching signals entirely.