Show HN: Energy-Efficient NAS via RBF Kernel Scoring (No GPU Training Needed)

1 point

a year ago

We present RBFleX-NAS, a training-free neural architecture search (NAS) framework that leverages an RBF kernel-based scoring mechanism to rank neural networks without training. Unlike traditional NAS methods that consume significant GPU resources, RBFleX-NAS can find high-performing architectures in seconds.

Key Applications: • Edge AI Deployment: Efficient model search for Raspberry Pi, Jetson, and mobile platforms • AutoML Integration: Lightweight NAS module suitable for AutoML systems • Cross-domain Transfer: Supports both vision and NLP tasks (e.g., TransNAS-Bench)

How it works: Instead of training each candidate network, we score architectures using a Radial Basis Function (RBF) kernel that captures both activation patterns and structural features.

Demo Video: https://youtu.be/QZz8s95x9xw?si=fqVs7T6no66zz_d5

Code on GitHub: https://github.com/tomomasayamasaki/RBFleX-NAS

Paper https://ieeexplore.ieee.org/document/10959729/metrics