SuperCM uses class-wise moving-average centroids to add a Gaussian-mixture clustering loss to a supervised classifier, improving CIFAR-10 SSL accuracy at low label counts.
Our training strategy benefits from the built-in cluster- ing capability of the CM module and does not rely on com- plex training schemes
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Supercm: Revisiting Clustering for Semi-Supervised Learning
SuperCM uses class-wise moving-average centroids to add a Gaussian-mixture clustering loss to a supervised classifier, improving CIFAR-10 SSL accuracy at low label counts.