A new attention-only white-box Transformer, AoT-ADMM, is derived via ADMM unrolling of a sparse rate reduction objective, reaches CRATE-comparable accuracy with 31% fewer parameters, and a ViT variant suggests MLP blocks may be largely redundant.
Advances in Neural Information Processing Systems , volume =
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Attention-Only White-Box Transformer via LeJEPA-Based Self-Supervised Pretraining
A new attention-only white-box Transformer, AoT-ADMM, is derived via ADMM unrolling of a sparse rate reduction objective, reaches CRATE-comparable accuracy with 31% fewer parameters, and a ViT variant suggests MLP blocks may be largely redundant.