Transformer -1 : Input-Adaptive Computation for Resource-Constrained Deployment

Ji Shihao, Song Zihui, Zhong Fucheng · 2025

Read Transformer -1 : Input-Adaptive Computation for Resource-Constrained Deployment by Ji Shihao, Song Zihui, Zhong Fucheng (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-16394.txt.

Description

Shihao Ji and co-authors give a transformer a throttle, spending compute only where an input deserves it. Fixed-depth networks waste effort on easy inputs and starve hard ones. Their design, Transformer inverse, pairs a complexity predictor with a learning-based policy that picks a computation path per input, trained end to end.

They back the scheme with a lower bound on how little dynamic computation can cost, plus engineering tricks: folding layers and pre-compiling graphs so the shape-shifting stays fast. On a standard image benchmark the method cut arithmetic by over forty percent and peak memory by about a third at matched accuracy. A trial on an edge chip and several language tasks suggest the savings travel.

Catalog record

CREATORJi Shihao, Song Zihui, Zhong Fucheng
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE31 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2501-16394

Browse the full library · About the Tower of Babel