FORGE: Forward-Only Test-Time Adaptation for Integer-Only Vision Models on Microcontrollers

Author1, Author2, Author3, Author4, Author5

test-time adaptationinteger-only modelsmicrocontrollerscomputer vision

2609.01683

Builder Relevance

80%

2h ago

Abstract

This paper presents a forward-only test-time adaptation method for integer-only vision models deployed on microcontrollers, addressing the challenges of adapting to distribution shifts without backpropagation.

Reality Card

Core Claim

The proposed method recovers most of the accuracy gains from gradient-based adaptation while being the only approach that operates on folded integer-only models.

Method / Result

Achieves 93% of the accuracy benefit by adapting only 3 of 21 layers.

Limitations

The method's reliance on specific layer adaptation may limit generalizability to other architectures or datasets.

Paper to code

Verified implementation resources so builders can test the paper’s claims instead of stopping at the abstract.

No verified implementation link has been attached yet. AIBuzzHub will keep this panel separate from unverified search results.

← Back to all papers