Beyond Empirical Support: Structured Outlier Generation via Sinkhorn Optimal Transport
The paper proposes a controlled way to generate rare stress cases outside a dataset’s observed support.
SBOG uses Sinkhorn optimal transport geometry to steer latent-space sampling toward weakly supported boundary regions. The authors say semantic constraints keep those samples tied to the intended context instead of drifting into arbitrary sparse areas. Experiments cover time-series anomaly generation and image outlier synthesis, where the method is reported to produce informative outliers for robustness evaluation. ArXiv · AI/CL/LG's note
SBOG uses Sinkhorn optimal transport geometry to steer latent-space sampling toward weakly supported boundary regions. The authors say semantic constraints keep those samples tied to the intended context instead of drifting into arbitrary sparse areas. Experiments cover time-series anomaly generation and image outlier synthesis, where the method is reported to produce informative outliers for robustness evaluation. ArXiv · AI/CL/LG's note
score 4