FeatureFormer predictor improves neural network performance estimation for edge devices

Researchers have developed FeatureFormer, a novel neural performance predictor designed to accurately estimate latency and energy consumption for neural networks on resource-constrained edge devices. This predictor incorporates explicit node-wise encodings of FLOPs, parameter counts, and memory proxies within a gated graph attention architecture. FeatureFormer demonstrates state-of-the-art performance on both latency and energy metrics, even in out-of-domain scenarios, and can improve existing predictors with minimal overhead. The research also introduces NNEQ, a new large-scale dataset for evaluating energy consumption. AI

IMPACT Enhances efficiency in neural architecture search for edge devices by improving performance prediction.

RANK_REASON Academic paper detailing a new model and dataset. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• alphaXiv
• arXiv
• CatalyzeX
• DagsHub
• FeatureFormer
• FLOPS
• Gotit.pub
• graph neural network
• Hugging Face
• NNEQ
• ScienceCast

• paper
• infra

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →