Gender recognition of human behaviors using neural ensembles

We have developed two ensembles of neural network classifiers in order to recognize actors' gender from their biological movements. One is the ensemble of modular MLPs (experts), the other is the ensemble of modular MLPs and an inductive decision tree which combines the output of experts. The human movement database consists of 13 males' and 13 females' movements, and contains 10 repetitions of knocking, waving and lifting movements both in neutral and angry style. Features have been extracted with 4 different representations such as the 2D and 3D velocities and positions, recorded from 6 point lights attached on body. We have compared the results of ensembles to the regular classifiers such as MLP, decision tree, self-organizing map and support vector machine. Furthermore, the discriminability and efficiency have been calculated for the comparison with the human performance that has been obtained with the same experiment. Our experimental results indicate that the ensemble models are superior to the conventional classifiers and human participants.