Deep learning for biometric face recognition: experimental study on benchmark data sets

There are still problems in applications of Machine Learning for face recognition. Such factors as lighting conditions, head rotations, emotions, and view angles affect the recognition accuracy. A large number of recognition subjects requires complex class boundaries. Deep Neural Networks have provided efficient solutions, although their implementations require massive computations for evaluation and minimisation of error functions. Gradient algorithms provide iterative minimisation of the error function. A maximal performance is achieved if parameters of gradient algorithms and neural network structures are properly set. The use of pairwise neural network structures often improves the performance because such structures require a small set of optimisation parameters. The experiments have been conducted on some face biometric benchmark data sets, and the main findings are presented in the form of a tutorial.