exemplar selection

**Exemplar selection** is the process of choosing **which specific examples to store** in a limited memory buffer for continual learning. Since buffer space is constrained, selecting the most informative, representative, or useful examples is critical for maximizing knowledge retention with minimal storage. **Selection Strategies** - **Random Selection**: Choose examples uniformly at random. Surprisingly effective and serves as a strong baseline. - **Herding (iCaRL)**: Select examples whose feature-space mean best approximates the overall class mean. Greedily picks the example that minimizes the distance between the buffer mean and the true class mean. - **K-Center Coreset**: Select examples that maximize **coverage** of the feature space — each selected example should represent a different region of the data distribution. - **Entropy-Based**: Select examples where the model is most **uncertain** (high entropy in predictions). These boundary examples are often most informative. - **Gradient-Based**: Select examples whose gradients are most representative of the overall gradient direction for the task. - **Diversity Maximization**: Select examples that are maximally different from each other, ensuring broad coverage. - **Reservoir Sampling**: Maintain a statistically uniform sample without needing to see all data at once — ideal for streaming settings. **Evaluation Criteria** - **Representativeness**: Do the selected examples capture the diversity and distribution of each class? - **Discriminativeness**: Do the selected examples preserve decision boundaries between classes? - **Compactness**: Can a small number of examples achieve performance close to replaying all data? **Task-Specific Considerations** - **Class-Balanced Selection**: Ensure each class has equal representation in the buffer — critical for maintaining balanced performance. - **Difficulty Balancing**: Store a mix of easy (typical) and hard (boundary) examples — easy examples for maintaining core knowledge, hard examples for preserving decision boundaries. - **Temporal Diversity**: For tasks with temporal patterns, select examples spanning the full time range rather than concentrating on one period. **Impact on Performance** The choice of exemplar selection strategy can affect continual learning accuracy by **3–10 percentage points** over random selection, with herding and coreset methods generally performing best. Exemplar selection is a **subtle but high-impact** design decision — the right selection strategy can dramatically improve knowledge retention within fixed memory constraints.