A naukri.com initiative
Arxiv
2d
46
Image Credit: Arxiv
Canonical Latent Representations in Conditional Diffusion Models
- Conditional diffusion models (CDMs) have shown impressive performance in generative tasks by modeling the full data distribution.
- CDMs can entangle class-defining features with irrelevant context, making it challenging to extract robust and interpretable representations.
- A new concept, Canonical Latent Representations (CLAReps), has been introduced to address this issue.
- CLAReps are latent codes in CDMs that preserve essential categorical information while discarding non-discriminative signals.
- By utilizing CLAReps, a novel diffusion-based feature-distillation paradigm called CaDistill has been developed.
- CaDistill ensures the transfer of core class knowledge from teacher to student CDMs via CLAReps.
- CLAReps enable representative sample generation for each class, providing an interpretable and compact summary of core class semantics.
- The student model trained with CaDistill achieves strong adversarial robustness and generalization ability.
- By focusing on class signals and ignoring spurious background cues, the student model becomes more robust.
- The study indicates that CDMs can serve not only as image generators but also as compact, interpretable teachers for driving robust representation learning.
Read Full Article
2 Likes
For uninterrupted reading, download the app