AI Paper Atlas

Research map / Architectures and efficiency / Neural architecture search

MCUFormer: Deploying Vision Tranformers on Microcontrollers with Limited Memory

Yinan Liang, Ziwei Wang, Xiuwei Xu, Yansong Tang, Jie Zhou, Jiwen Lu

NeurIPS 2023 · Architectures and efficiency › Neural architecture search · approach: Latency-aware architecture search · 3 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use proposed search space evolution instead of random search space (ablation)MCUFormer · ImageNettop-1 accuracy: 65.2% -> 71.1%Improved
use proposed search space evolution instead of maximal search space (ablation)MCUFormer · ImageNettop-1 accuracy: 69.2% -> 71.1%Improved
use proposed evolution instead of composing search space selection into supernet training (ablation)MCUFormer · ImageNettop-1 accuracy: 68.6% -> 71.1%Improved
neighborhood size 2 vs 5 in tendency fitting (ablation)MCUFormer · ImageNettop-1 accuracy: 71.1% -> 65.4%Hurt
neighborhood size 3 vs 5 in tendency fitting (ablation)MCUFormer · ImageNettop-1 accuracy: 71.1% -> 68.2%Hurt
neighborhood size 8 vs 5 in tendency fitting (ablation)MCUFormer · ImageNettop-1 accuracy: 71.1% -> 71.1%No clear effect
add operator integration (int8 operators) to operator library (ablation)MCUFormer operator library · peak memorypeak memory: 23.8%Improved
add patch embedding decomposition, operator integration and token overwriting together (ablation)MCUFormer operator library · peak memorypeak memory: 9%Improved
medium update thresholds h1 and h2 vs large or small (ablation)MCUFormer search space evolution · ImageNettop-1 accuracyMixed
replace AutoFormer search with proposed search space evolution (ablation)MCUFormer-A vs MCUFormer · ImageNet STM32F7top-1 accuracy: 69.3% -> 73.6%Improved
replace TinyEngine with proposed operator library (ablation)MCUFormer-T vs MCUFormer · ImageNet STM32F7peak memory: 872kB -> 319kBImproved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy