Combining diffusion-generated ImageNet-100 clones with synthetic hard negatives lifts DeiT-S linear probing to 82.12, but only with double the training iterations; at matched compute the synthetic components do not help DeiT-S and barely help Swin-T.
Unsupervised learn- ing of visual features by contrasting cluster assignments
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
citation-role summary
background 1
citation-polarity summary
fields
cs.CV 1years
2025 1verdicts
CONDITIONAL 1roles
background 1polarities
background 1representative citing papers
citing papers explorer
-
Fake & Square: Training Self-Supervised Vision Transformers with Synthetic Data and Synthetic Hard Negatives
Combining diffusion-generated ImageNet-100 clones with synthetic hard negatives lifts DeiT-S linear probing to 82.12, but only with double the training iterations; at matched compute the synthetic components do not help DeiT-S and barely help Swin-T.