task-incremental learning

**Task-Incremental Learning** is a **continual learning paradigm where a model sequentially acquires new tasks while retaining performance on previously learned ones, with the critical advantage that task identity is provided at test time** — enabling multi-head neural architectures to achieve near-zero catastrophic forgetting by routing inputs to task-specific output layers, while still requiring regularization or memory replay to prevent degradation of shared lower-level representations. **What Is Task-Incremental Learning?** - **Definition**: A continual learning setting where tasks arrive sequentially (T1, T2, ..., TN), each with distinct label spaces or objectives, and the model receives the task identifier at test time to select the appropriate prediction head. - **Task Oracle Assumption**: Unlike class-incremental or domain-incremental learning, task-incremental learning assumes task label is known at inference — significantly simplifying the problem. - **Multi-Head Architecture**: Each task typically receives its own output layer (head) while sharing lower-level feature representations, enabling task-specific prediction without cross-task label confusion. - **Sequential Training**: Tasks are learned one at a time without revisiting previous task data (or with limited replay buffers), simulating real-world lifelong learning constraints. **Why Task-Incremental Learning Matters** - **Lifelong AI Systems**: Real-world AI assistants must accumulate new skills (languages, domains, tasks) without forgetting existing capabilities. - **Privacy-Preserving Learning**: Organizations cannot always store historical data due to GDPR and retention policies — sequential learning must work without data revisitation. - **Reduced Catastrophic Forgetting**: The task identity oracle makes this the most tractable continual learning setting, enabling near-perfect retention with proper architecture. - **Foundation for Harder Settings**: Understanding task-incremental dynamics informs solutions for harder variants (class-incremental, domain-incremental). - **Benchmark Clarity**: Standardized evaluation protocols (permuted MNIST, Split CIFAR-100) enable rigorous comparison of continual learning algorithms. **Approaches to Task-Incremental Learning** **Architectural Methods**: - **Progressive Networks**: Add new network columns per task; freeze previous columns with lateral connections — zero forgetting but linear memory growth. - **PackNet**: Iterative pruning frees network capacity for new tasks while retaining compressed representations of old ones. - **Dynamic Expandable Networks**: Selectively expand model capacity based on task novelty — balances growth and reuse. **Regularization Methods**: - **EWC (Elastic Weight Consolidation)**: Penalizes changes to weights important for previous tasks using Fisher Information as importance metric. - **SI (Synaptic Intelligence)**: Online estimation of weight importance during training — no post-hoc Fisher computation required. - **LwF (Learning without Forgetting)**: Knowledge distillation from old task predictions to preserve behavior without storing old data. **Replay Methods**: - **Experience Replay**: Store small buffer of previous task examples; interleave with new task training to prevent forgetting. - **Generative Replay**: Use generative model to synthesize pseudo-examples from past tasks — eliminates storage of real data. - **Dark Experience Replay (DER)**: Replay stored logits rather than raw inputs — more storage-efficient and better preserves decision boundaries. **Method Comparison** | Method | Forgetting | Memory | Compute | |--------|-----------|--------|---------| | Multi-head only | High (shared layers) | Low | Low | | EWC | Medium | Low | Medium | | Experience Replay | Low | Medium | Medium | | Progressive Nets | Zero | High | High | Task-Incremental Learning is **the foundation of lifelong machine intelligence** — providing the theoretical and algorithmic basis for AI systems that accumulate knowledge continuously, retaining past expertise while growing into new domains without catastrophic forgetting of previously mastered tasks.