On the choice of the non-trainable internal weights in random feature maps