Suppose you have a Model m that outputs an embedding vector e. In the normal binary classification task, e is transformed into a score s by multiplying with weight matrix W, which has a single row vector. The score s is then passed through the sigmoid activation function and the resulting value (p) is treated as a probability and used in a binary-cross-entropy loss function. In a new setting, the weight matrix W will have 4 row vectors, resulting in 4 scores (s1 - s4 ). The probability value p is calculated as follows: p = sigmoid ( max (s1, s2, s3, s4 ) )