Understanding Logits, Sigmoid, Softmax, and Cross-Entropy Loss in Deep Learning