We introduce GHGR, a robust neural estimator of the Hirschfeld-Gebelein-Rényi (HGR) maximal correlation, which efficiently detects nonlinear dependencies, satisfies Rényi's axioms, and outperforms state-of-the-art methods.