Affinity Graph Supervision for Visual Recognition

Graphs as Tools to Improve Deep Learning Methods

“…In this subsection, we present another way of defining the loss of a DNN using graphs [59]. During training, a DNN is evaluated on a whole batch of data samples before being updated.…”

“…During training, a DNN is evaluated on a whole batch of data samples before being updated. The idea described in [59] is to ease the training by exploiting meaningful relationships between these data samples. In short, the idea amounts to add a regularization term to the usual loss of the DNN.…”

“…The affinity between two samples x i and x j within the DNN can be computed with any similarity function W, as Now, the idea in [59] is to select a meaningful set of connections that we will aim at maximizing during training. For instance, in a classification task, we want to maximize the similarity between the outputs of data samples of the same class.…”

“…In [59], the maximization of M is expressed as the minimization of a loss L G . A discussion on the choice of the loss form is made in the original paper [59].…”

“…In [59], experiments are made on CIFAR-10, CIFAR-100 and Tiny-ImageNet are summarized. The experiments show that there is an improvement in terms of classification accuracy over a baseline, when using the affinity-based loss.…”