OnEmpiricalRiskMinimizationwithDependent and Heavy-TailedData