relational knowledge distillation

**Relational Knowledge Distillation (RKD)** is a **distillation method that transfers the geometric relationships between samples rather than individual sample representations** — teaching the student to preserve the distance and angle structure of the teacher's feature space. **How Does RKD Work?** - **Distance-Wise**: Minimize $sum_{(i,j)} l(psi_D^T(x_i, x_j), psi_D^S(x_i, x_j))$ where $psi_D$ is the pairwise distance function. - **Angle-Wise**: Preserve the angle formed by triplets of points in the embedding space. - **Representation**: Instead of matching individual features, match the relational structure (distances, angles) between sample pairs/triplets. - **Paper**: Park et al. (2019). **Why It Matters** - **Structural Knowledge**: Captures the manifold structure of the feature space, not just point-wise values. - **Robustness**: Less sensitive to absolute scale differences between teacher and student representations. - **Metric Learning**: Particularly effective for tasks where relative distances matter (face recognition, retrieval). **RKD** is **transferring the geometry of knowledge** — teaching the student to arrange its representations in the same relative structure as the teacher, regardless of absolute coordinates.