Learning from a tiny dataset of manual annotations: a teacher/student approach for surgical phase recognition