Continual Learning and Catastrophic Forgetting is the fundamental challenge in neural network training where a model trained sequentially on multiple tasks loses performance on earlier tasks as it adapts to new ones — because gradient-based updates to accommodate new data overwrite the weight configurations that encoded previous knowledge, requiring specialized techniques (EWC, progressive networks, replay) to maintain performance across all tasks without access to previous training data.

The Catastrophic Forgetting Problem

When a model trained on Task A is subsequently trained on Task B, its performance on Task A degrades dramatically — often to random chance. This happens because the loss landscape for Task B pulls weights away from the region optimal for Task A. Standard SGD has no mechanism to preserve previously learned representations. This is fundamentally different from human learning, where acquiring new skills enhances rather than overwrites existing knowledge.

Continual Learning Strategies

Regularization-Based Methods:

• EWC (Elastic Weight Consolidation): Identifies which weights are most important for previous tasks using the Fisher Information Matrix (diagonal approximation). A quadratic penalty discourages changes to important weights: L_total = L_new + λ × Σ F_i × (θ_i - θ*_i)². Important weights are "elastic" — resistant to change.
• SI (Synaptic Intelligence): Computes weight importance online during training by tracking the contribution of each weight to the loss decrease. No need for a separate Fisher computation step.
• Learning without Forgetting (LwF): Uses knowledge distillation — the model's predictions on new task data (before training) serve as soft targets that the model must continue to match after training on the new task.

Replay-Based Methods:

• Experience Replay: Store a small buffer of examples from previous tasks. Interleave buffer samples with new task data during training. Simple and effective but requires storing raw data (privacy concerns).
• Generative Replay: Train a generative model (VAE, GAN) on previous task data. Generate synthetic examples from previous tasks to mix with new data. No raw data storage needed.
• Dark Experience Replay: Store model logits (soft predictions) alongside raw examples. Replay both data and the model's previous response to that data.

Architecture-Based Methods:

• Progressive Neural Networks: Add new columns (sub-networks) for each task with lateral connections to previous columns. Previous columns are frozen — zero forgetting by construction. Disadvantage: model grows linearly with number of tasks.
• PackNet: Prune the network after each task (identify important weights, freeze them). Remaining free weights are available for the next task. Model capacity is gradually consumed.
• Adapter Modules: Add small task-specific adapter layers while keeping the backbone frozen. Each task gets its own adapters. Similar to multi-LoRA serving for LLMs.

Evaluation Protocol

• Average Accuracy: Mean accuracy across all tasks after training on the final task.
• Backward Transfer (BWT): Average change in performance on previous tasks after training new ones. Negative BWT = forgetting.
• Forward Transfer (FWT): Influence of previous task training on performance on new tasks before training on them.

Continual Learning is the unsolved grand challenge of making neural networks learn like humans — accumulating knowledge over time without forgetting, a capability that would transform AI from systems that are trained once to systems that grow continuously more capable through experience.