Catastrophic Forgetting

The fundamental trade-off at the heart of catastrophic forgetting. A network must possess plasticity to encode new information by updating its weights, yet it requires stability to retain existing knowledge. Standard gradient descent lacks a native mechanism to protect weights critical for prior tasks. When a new data distribution is introduced, the loss landscape shifts, and the optimizer follows the steepest descent path, which often overwrites the weights that encoded previous tasks. This is not a bug but an inherent consequence of using a single, shared representational space for sequential learning.

Forgetting severity is directly proportional to the degree of representational overlap between sequential tasks. If Task A and Task B rely on similar features in hidden layers, learning Task B will shift those shared features to optimize for the new objective, destroying their utility for Task A. This is visualized as a destructive interference pattern in the weight space. Conversely, tasks with highly orthogonal representations exhibit less interference. This characteristic explains why fine-tuning a general-purpose model on a narrow domain often degrades its broad capabilities.

Unlike gradual skill decay in biological systems, catastrophic forgetting in artificial neural networks manifests as an abrupt phase transition. Performance on a previous task does not degrade linearly. Instead, it remains stable for a number of training steps on the new task before collapsing sharply. This 'cliff-edge' behavior occurs because the optimizer remains in a local minimum for the old task until a weight update pushes it out of that basin of attraction. This makes early stopping an unreliable mitigation strategy.

The network's final weights become heavily biased toward the most recently seen data distribution. This recency bias is a direct result of the sequential nature of training. The final gradient updates overwrite the decision boundaries learned from earlier tasks, effectively erasing the model's ability to discriminate classes from the initial dataset. In a classification setting, the model will often predict all inputs as belonging to classes from the most recent task, demonstrating a complete loss of earlier class distinctions.

The rate of forgetting is highly sensitive to the statistical divergence between the old and new input distributions. A large distributional shift, such as moving from natural images to medical scans, triggers more aggressive weight adaptation and thus more severe forgetting. Even subtle shifts, like a change in lighting conditions or background texture, can induce measurable forgetting if the network has overfit to spurious correlations in the original training set. This characteristic directly links catastrophic forgetting to the broader challenge of domain generalization.

In deep networks, catastrophic forgetting exhibits a layer-specific signature. Early layers, which learn general-purpose feature detectors like edge and texture filters, tend to stabilize quickly and are less susceptible to overwriting. The most severe forgetting occurs in the higher, task-specific layers where representations are more abstract and semantically entangled. This observation underpins mitigation strategies like Elastic Weight Consolidation (EWC), which selectively constrains plasticity in the most critical weights for previous tasks.

A synaptic consolidation algorithm that identifies parameters critical to previously learned tasks and penalizes their modification during new training. EWC calculates the Fisher Information Matrix to approximate the importance of each weight, applying a quadratic penalty proportional to this importance. This constrains plasticity in high-importance synapses while allowing low-importance weights to adapt freely.

• Mechanism: Adds a regularization term to the loss function that anchors critical weights.
• Key Insight: Mimics the biological synaptic consolidation observed in mammalian brains.
• Limitation: Computational cost scales quadratically with the number of parameters; requires storing a separate Fisher matrix per task.

An architectural approach that instantiates a new neural network column for each new task while freezing previously learned columns. Lateral connections from frozen columns to the new column enable forward transfer of features without risking overwriting.

• Mechanism: Blocks are added laterally, preserving prior weights immutably.
• Advantage: Guarantees zero forgetting by design.
• Trade-off: Parameter count grows linearly with the number of tasks, making it impractical for lifelong learning scenarios with hundreds of tasks.

A data-centric strategy that interleaves samples from previous tasks into the training stream for the new task. By maintaining a replay buffer of stored exemplars, the model continuously rehearses old knowledge, preventing its decision boundaries from being overwritten.

• Variants: Exact replay (storing raw data) vs. generative replay (training a generative model to synthesize past samples).
• Constraint: Exact replay violates data retention policies in privacy-sensitive domains; generative replay can suffer from mode collapse.

A distillation-based method that uses the model's own predictions on old tasks as soft targets while learning a new task. The network is trained to minimize both the cross-entropy loss for the new task and a distillation loss that encourages the new model's output distribution to match the frozen copy's output on old task data.

• Mechanism: Knowledge distillation from a frozen snapshot of the model.
• Benefit: Requires no access to old training data.
• Risk: Distillation loss may not perfectly preserve fine-grained decision boundaries.

An importance-weighting method that estimates parameter importance based on the sensitivity of the learned function's output to weight perturbations, rather than relying on the loss gradient. MAS accumulates importance by measuring the magnitude of the gradient of the squared L2-norm of the model's output with respect to each parameter.

• Advantage over EWC: Operates in an unsupervised manner, requiring no labeled data to compute importance.
• Use Case: Effective for continual learning in reinforcement learning environments where loss signals are noisy.

A network pruning approach that iteratively trains on sequential tasks and then prunes redundant parameters, freeing capacity for subsequent tasks. After pruning, the surviving weights are frozen, and the pruned connections are re-initialized for the next task.

• Mechanism: Iterative magnitude-based pruning to create task-specific sparse sub-networks.
• Result: A single compact model with dedicated, non-overlapping parameter sets for each task.
• Constraint: Requires knowing the total number of tasks in advance to allocate the pruning budget correctly.