← AI Techniques Radar
Training & fine-tuning

Mode-Aware Prototype Allocation (MAPA) for Latent Dataset Distillation

A latent-space dataset distillation approach for 3D medical classification that allocates synthetic prototypes across intra-class latent modes to prevent “prototype concentration” in only dense regions, while still maintaining class-level distribution matching.

Discovered 2026-11-10 · Sources checked 2026-11-10

This summary was generated with AI assistance.

Why it matters

In latent distillation, class-level matching can unintentionally concentrate prototypes in dense latent areas and underrepresent rarer intra-class modes. This can reduce coverage of clinically relevant variability within a class. MAPA explicitly targets mode coverage within each class to improve balanced intra-class performance.

What changed

Introduces (1) fixed mode anchors per class obtained by clustering real latent representations, (2) a soft assignment mechanism that spreads prototypes across these anchors, (3) a lightweight anti-collapse regularizer to avoid prototype collapse, and (4) curriculum weighting to stabilize optimization.

Where it helps

  • Latent dataset distillation for 3D medical image classification
  • Prototype-based synthesis where class-level matching causes intra-class mode undercoverage
  • Balanced performance optimization across heterogeneous intra-class patterns

Evidence & reproduction

Reproduction not evaluated. A linked paper or repository does not establish independent reproduction.

Limitations & open questions

The described method is presented in the context of 3D medical classification using a frozen autoencoder latent space; details about applicability to other modalities or architectures are not specified.

Sources