Learning to Optimize with Unsupervised Learning: Training Deep Neural Networks for URLLC