continual

**Continual Learning Catastrophic Forgetting** is **training neural networks sequentially on tasks without forgetting previously learned tasks, addressing the catastrophic forgetting problem where new learning overwrites old knowledge** — enabling lifelong AI systems. Continual learning mimics human learning. **Catastrophic Forgetting** neural networks trained on sequence of tasks forget earlier tasks. Weights optimized for task 2 become poor for task 1. Plasticity-stability dilemma: adapt to new tasks (plasticity) while maintaining old knowledge (stability). **Task Incremental Learning** tasks arrive sequentially. Network must: learn current task, remember previous tasks. Task identity available at test time (task-specific output head). **Class Incremental Learning** new classes arrive over time. No task boundaries. Single output head. More difficult than task incremental. **Domain Incremental Learning** same task, data distribution changes. Covariate shift between tasks. **Replay and Experience Replay** remember subset of old data, replay during new task training. Interleave old and new task. Effective but requires storing past data. **Generative Replay** generate pseudo-examples of old tasks via generative model. No storage of real data but generative model adds complexity. **Elastic Weight Consolidation (EWC)** track importance of weights for previous tasks via Fisher information matrix. Penalize changes to important weights: loss = new_loss + λ * Σ F_i * (w_i - w_i*)^2. F_i = Fisher information (importance). **Synaptic Importance** different parameterizations of importance. Elastic weight consolidation, synaptic importance, MAS (Memory Aware Synapses). **Memory Consolidation** biological inspiration: brain consolidates memories during sleep. Offline consolidation phase after task. **Dynamic Expansion** add new neurons for new tasks. Gradually increase capacity. Avoid catastrophic forgetting through architecture expansion. **PackNet** mask learning: learn binary masks per task indicating which weights to use. Enables selective reuse. **Adapter Modules** small trainable modules for each task. Keep base network frozen. Task-specific learning through adapters. **Prompt Learning** condition network on task-specific prompts. Learn prompts, reuse backbone. Similar to adapter idea. **Domain-Aware Learning** use domain information to guide consolidation. Separate out domain-specific and task-specific factors. **Sparse Representations** sparse activations naturally avoid interference. Active neurons for task 1 different from task 2. **Disentangled Representations** learn separated representations for different factors. Disentanglement reduces interference. **Backward Transfer** learning new task improves old task performance. Positive: generalization. **Forward Transfer** learning old task helps new task. Domain overlap and transfer. **Meta-Learning for Continual Learning** learn learning algorithm that avoids forgetting. MAML, other meta-learning approaches. **Rehearsal-Free Methods** don't replay old data. Replay impractical at scale. **Pseudo-Rehearsal** synthetic examples of old tasks. Generate via generative model. **Curriculum Learning** order tasks for efficient learning. Easier tasks first. Smooth transition between tasks. **Evaluation Metrics** final accuracy on all tasks, backward transfer, forward transfer. Metrics differ from standard supervised learning. **Benchmarks** Permuted MNIST, Split CIFAR-10/100, ImageNet-100. **Biological Plausibility** brain continually learns. Synaptic consolidation, neuromodulation mechanisms. **Practical Challenges** computational efficiency (repeated learning slows down). Scalability (many tasks). **Applications** robots learning sequentially, dialogue systems improving over time, personalized ML systems. **Continual learning enables AI systems learning throughout deployment** rather than static pretrained models.