Catastrophic Forgetting

This is the fundamental trade-off at the heart of catastrophic forgetting. A network must possess plasticity to integrate new knowledge from a support set, but it requires stability to retain previously learned representations. Without a mechanism to balance these forces, new weight updates overwrite the parameters critical for old tasks. The network's embedding space shifts entirely to accommodate the new query set distribution, destroying the decision boundaries that defined prior classes.

Forgetting occurs because gradient descent optimizes the entire parameter space for the current loss landscape without regard for historical importance. When a model trained on Task A is fine-tuned on Task B, the optimizer unapologetically moves weights away from the Task A minimum to find the Task B minimum. Key mechanisms include:

• Representation Drift: Feature extractors lose sensitivity to previously critical patterns.
• Output Interference: New classification heads cause logit suppression for old classes.
• Unconstrained Plasticity: All weights are equally eligible for modification, destroying the fragile parameter configurations that encoded prior knowledge.

A seminal synaptic consolidation algorithm that acts as a computational proxy for biological memory protection. EWC calculates the Fisher Information Matrix to identify which weights are crucial for a previously learned task. It then applies a quadratic penalty to changes on those specific parameters during subsequent training. This selectively slows learning on high-importance weights, forcing the optimizer to find a solution for the new task that resides in a low-importance parameter subspace, thereby preserving old knowledge.

A rehearsal-based approach that interleaves samples from previous tasks into the training stream for the new task. By maintaining a small episodic memory buffer of raw inputs or generating pseudo-samples via a generative model, the network is constantly reminded of its past. This prevents the data distribution from shifting entirely to the new domain. The joint optimization on mixed batches maintains the separation of old and new classes in the embedding space, directly countering the catastrophic shift in decision boundaries.

A family of techniques that combat forgetting by allocating distinct, non-overlapping neural resources to different tasks. Methods include:

• Progressive Networks: Freeze old task columns and instantiate new lateral connections.
• Dynamic Architecture Expansion: Add new neurons or layers for new tasks while masking gradients from old parameters.
• PathNet: Uses genetic algorithms to find distinct pathways through a fixed, large network for each task. This hard separation guarantees zero interference, achieving perfect stability at the cost of scalability.

In RF fingerprinting, catastrophic forgetting is a critical operational failure mode. If a model trained to authenticate 100 known transmitters undergoes few-shot enrollment to add a single new device, it risks forgetting the signatures of the original 100. This manifests as a spike in the False Rejection Rate (FRR) for the legacy fleet. Mitigation via EWC or episodic memory ensures that rapid onboarding of new IoT hardware does not invalidate the identity models of the existing deployed infrastructure.

A seminal algorithm that mitigates catastrophic forgetting by identifying and protecting the synaptic weights most critical to previously learned tasks. It works by approximating the Fisher Information Matrix to calculate the importance of each weight. When learning a new task, EWC applies a quadratic penalty that selectively slows down learning on important parameters, effectively anchoring old knowledge while allowing less critical weights to adapt. This is analogous to a spring mechanism, where stiffer springs represent higher importance.

The fundamental trade-off at the heart of catastrophic forgetting. Stability is the ability of a network to retain existing knowledge, while plasticity is its capacity to integrate new information. A perfectly stable network cannot learn anything new, while a perfectly plastic network forgets everything instantly. The goal of continual learning is to find an optimal balance, creating artificial systems that are plastic enough to acquire new skills but stable enough to not overwrite old ones.

A biologically inspired rehearsal technique where the network retains and interleaves a small subset of samples from previous tasks during training on a new task. By periodically replaying old data, the model's weights are constrained to a region of the parameter space that performs well on all tasks. This directly combats catastrophic forgetting by preventing the data distribution from shifting entirely to the new domain. Key variants include:

• Exact replay: Storing raw samples in a memory buffer.
• Generative replay: Training a generative model to synthesize pseudo-samples of old tasks.

An algorithm that, unlike EWC, computes weight importance online and in an unsupervised manner during the training trajectory. SI tracks the contribution of each individual synapse to the total reduction in loss over time. When a new task is introduced, a regularizer penalizes changes to synapses that were major contributors to previous learning. This approach is task-agnostic and does not require a separate importance calculation phase at the end of a task, making it suitable for continuous, online learning scenarios.

A specific manifestation of catastrophic forgetting where a model's predictions become heavily skewed toward the most recently learned task. This occurs because the final-layer representations and classification heads are overwritten by new data. In a sequential learning setting, a model suffering from task-recency bias will often classify all inputs as belonging to the latest task, regardless of their true origin. Mitigation often involves architectural strategies like task-specific output heads or bias-correction layers.