Multi-task learning on the edge for effective gender, age, ethnicity and emotion recognition