NAISS
SUPR
NAISS Projects
SUPR
Layer-wise Quantized EuroSAT ViT, MLP-Mixer, ResNet-50, and others
Dnr:

NAISS 2026/4-1370

Type:

NAISS Small

Principal Investigator:

Pedro De Melo Antunes

Affiliation:

Kungliga Tekniska högskolan

Start Date:

2026-08-10

End Date:

2027-06-01

Primary Classification:

10210: Artificial Intelligence

Webpage:

Allocation

Abstract

Deploying state-of-the-art vision models, such as Vision Transformers (ViT), MLP-Mixers, and ResNet-50, on resource-constrained edge devices remains a significant challenge due to their high memory footprint and computational requirements. This project aims to address this by investigating layer-wise quantization strategies for these architectures, with a specific focus on remote sensing applications using the EuroSAT dataset. By employing Quantization-Aware Training (QAT) via the Brevitas library, we will explore how mixed-precision quantization can compress these models while preserving their predictive accuracy. The outcome of this research will be highly optimized, low-bitwidth models exported in QONNX format, paving the way for efficient, hardware-accelerated Earth observation systems on edge devices.