In the online lecture, Kayvon mentions that we could have biasing due to incorrect scheduling. Could someone explain what biasing in this context means? Does he mean bias in parameters or update rates or bias in the final network classifier?

Can it be reduced? If so, how?