This new paradigm of training a neural network for distance-based predictions instead of classification-based predictions requires a new loss function. Recall that in previous chapters, we used simple loss functions such as categorical cross-entropy to measure the accuracy of our predictions in classification problems.
In distance-based predictions, loss functions based on accuracy would not work. Therefore, we require a new distance-based loss function to train our Siamese neural network for facial recognition. The distance-based loss function that we will be using is called the contrastive loss function.
Take a look at the following variables:
- Ytrue: Let Ytrue be 1 if the two input images are from the same subject (same face) and 0 if the two input images are from different subjects (different faces)
- D: The predicted distance output from the neural network ...