Solving the same-different task with convolutional neural networks

Deep learning demonstrated major abilities in solving many kinds of different real-world problems computer vision literature. However, they are still strained by simple reasoning tasks that humans consider easy to solve. In this work, we probe current state-of-the-art convolutional neural networks on a difficult set known as the same-different problems. All require same prerequisite be solved correctly: understanding if two random shapes inside image or not. With experiments carried out demonstrate residual connections, and more generally skip seem have only marginal impact proposed particular, experiment with DenseNets, examine contribution recurrent connections already tested architectures, ResNet-18, CorNet-S respectively. Our show older feed-forward networks, AlexNet VGG, almost unable learn problems, except some specific scenarios. We recently introduced architectures can converge even cases where important parts their architecture removed. finally carry zero-shot generalization tests, discover these scenarios stronger overall test accuracy. On four from SVRT dataset, reach results respect previous approaches, obtaining super-human performances three