On the Generalization Properties of Learning the Random Feature Models with Learnable Activation Functions