We demonstrate a machine-learning-based traffic-aware approach for dynamic network slicing in optical networks. Experimental results indicate that the proposed framework achieves 96% traffic prediction accuracy, 49% blocking reduction and 29% delay reduction compared with conventional solutions.