CNN meets statistics on a single CPU

🔥 What happened Researchers fused a tiny CNN with a classical cumulant classifier for automatic modulation classification. Their FuseAMC model hits 55.8% accuracy on RadioML 2016.10a – 5.1 points above the best classical baseline and 3.2 above the CNN alone. 💡 Why it matters The CNN shines at low SNR, the cumulant classifier at high SNR – they're complementary. With just 0.2M parameters and 3 MFLOPs per frame, the whole pipeline trains and runs on a single CPU core, no hardware acceleration needed. ⚡ Our take Instead of chasing ever-larger models, this work shows how 200k parameters plus some statistical savvy can beat established deep-learning baselines. A wake-up call for anyone blindly betting on GPUs.