A dynamic and flexible generalized spatial modulation (GSM) framework is proposed for large-scale MIMO systems. Our framework is leveraged on the utilization of machine learning methods for GSM in order to improve the error performance in the presence of time-correlated channels and channel estimation errors. The decision tree and multi-layer perceptron algorithms are adopted as transmit antenna selection approaches. Simulation results indicate that in the presence of real-life impairments, machine learning based approaches provide a superior performance when compared to the classical Euclidean distance based approach. The observations are validated through measurement results over the designed 16 x 4 MIMO test-bed using software defined radio nodes.