In this letter, we propose a real-time self-collision avoidance approach for whole-body humanoid robot control. To achieve this, we learn the feasible regions of control in the humanoid's joint space as smooth self-collision boundary functions. Collision-free motions are generated online by treating the learned boundary functions as constraints in a Quadratic Program based Inverse Kinematic solver. As the geometrical complexity of a humanoid robot joint space grows with the number of degrees-of-freedom (DoF), learning computationally efficient and accurate boundary functions is challenging. We address this by partitioning the robot model into multiple lower-dimensional submodels. We compare performance of several state-of-the-art machine learning techniques to learn such boundary functions. Our approach is validated on the 29-DoF iCub humanoid robot, demonstrating highly accurate real-time self-collision avoidance.