minimize KL divergence
        ↓
minimize cross-entropy
        ↓
minimize negative log-likelihood