Input-dependent top-k sparse attention makes small transformers converge faster and generalize as well as full attention, while input-agnostic sparsity does not, and the effect is tied to reduced dispersion of attention scores.
The Senses: A Comprehensive Reference
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
-
Transformers Learn Faster with Semantic Focus
Input-dependent top-k sparse attention makes small transformers converge faster and generalize as well as full attention, while input-agnostic sparsity does not, and the effect is tied to reduced dispersion of attention scores.