Forming Auxiliary High-confident Instance-level Loss to Promote Learning from Label Proportions