Learning from Distributions via Support Measure Machines