Saved in:
Bibliographic Details
Main Authors: Hu, Jerry Yao-Chieh, Lu, Mingcheng, Lee, Yi-Chen, Liu, Han
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2604.24878
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866918470680051712
author Hu, Jerry Yao-Chieh
Lu, Mingcheng
Lee, Yi-Chen
Liu, Han
author_facet Hu, Jerry Yao-Chieh
Lu, Mingcheng
Lee, Yi-Chen
Liu, Han
contents We provide a systematic recipe for translating ReLU approximation results to softmax attention mechanism. This recipe covers many common approximation targets. Importantly, it yields target-specific, economic resource bounds beyond universal approximation statements. We showcase the recipe on multiplication, reciprocal computation, and min/max primitives. These results provide new analytical tools for analyzing softmax transformer models.
format Preprint
id arxiv_https___arxiv_org_abs_2604_24878
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Transformer Approximations from ReLUs
Hu, Jerry Yao-Chieh
Lu, Mingcheng
Lee, Yi-Chen
Liu, Han
Machine Learning
Artificial Intelligence
We provide a systematic recipe for translating ReLU approximation results to softmax attention mechanism. This recipe covers many common approximation targets. Importantly, it yields target-specific, economic resource bounds beyond universal approximation statements. We showcase the recipe on multiplication, reciprocal computation, and min/max primitives. These results provide new analytical tools for analyzing softmax transformer models.
title Transformer Approximations from ReLUs
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2604.24878