The following is taken from a paper
"Since a human can simultaneously perform multiple actions (e.g., sit and drink), our output layer consists of binary sigmoid classifiers for multilabel action classification (i.e. the predicted action classes do not compete). The training objective is to minimize the binary cross entropy losses between the ground-truth action labels and the scores predicted by the model.
Dear Tahmid Chowdhury ,
Great discussion point. Following the updates!
According to your limited descriptions, it seems that the output of final layer (sigmoid classifier) is a binary vector. Each bit of the binary vector is valued 0 or 1, which is the output of one binary sigmoid classifier. The type of output is also regarded as the one-hot coding for multiple labels.
Dear Tahmid Chowdhury ,
Great discussion point. Following the updates!
By using One-Class attributes versus all rest class attributes approach, considering the attribute of the one class as positives and remaining as negatives in training and the trained pattern is available to each one of the binary classifiers and then, the supervised algorithm which gives a positive output reveals the output class in validation phase ( From Literature Survey)
- Badges
- Science topic
- Similar topics
- Data Mining
- Classification