Probabilistic loss function for training network with triplets

A non-transitory machine-readable medium storing a program which when executed by at least one processing unit trains a machine-trained (MT) network that processes images using a plurality of network parameters, the program comprising sets of instructions for:

• propagating a triplet of input images through the MT network to generate an output vector for each of the input images, wherein the triplet comprises an anchor first image, a second image of a same category as the anchor first image, and a third image of a different category than the anchor first image;
• calculating a value of a loss function for the triplet that is a logarithm of the inverse of a probability that the output vector for the anchor image is classified as the same category as the second image rather than the same category as the third image based on an assumption that (i) a distribution of vectors for each of the categories is a normal distribution and (ii) a variance of the normal distribution is the same for each of the categories;
• and
• using the calculated loss function value for the triplet to train the network parameters by (i) backpropagating the calculated loss function value through the MT network to determine, for each network parameter of a set of the network parameters, a rate of change in the calculated loss function value relative to a rate of change of the network parameter, and (ii) modifying each network parameter in the set of network parameters based on the determined rate of change for the network parameter,
• wherein the MT network with trained network parameters is for embedding into a device to process images.