Computer Vision News - February 2023

5 TAME: Attention Mechanism Based... is unique. It’s structured like an attention mechanism but isn’t used the same way as a typical attention mechanism in the network, where it would usually go between network layers to be accessed at every forward pass. Instead, it is a branch to the original network that learns some form of weights to modify the original image at the training stage. ” A challenge TAME encountered was that the masks used a cross-entropy loss to learn the additional attention layer, but this loss was deriving wide masks that covered most of the area of the image to keep the confidence score high. The team needed to restrict the explanation mask produced to look at only the necessary pixels of the image. To solve this, they extended the loss with two additional parts: a variation loss to penalize the fragmentation of the image and focus on the most important small parts and an area loss to activate smaller recognize something . ” TAME uses computed weights of the feature maps from convolutional layers to derive an explanation mask and is unique in that it uses a training set to learn the optimal weight. Previously, methods used only one image to compute these weights at the inference stage, whereas TAME extracts knowledge from the abundance of training data that exists in the literature. After training, explanation maps can be computed with just a single pass. TAME’s two main novelties are how it formulates the loss function for performing the training and combines feature maps from different layers of the original DCNN classifier to generate an explanation map. “ We introduced an attention mechanism for learning from these feature maps, ” Vasileios explains. “ The way that’s used both at the training and inference stages