Abstract
We study the problem of processing structured data such as images or graphs by deep neural networks. Given a set of features extracted by convolutional or recurrent network, a global pooling is commonly applied to produce a fixed-length representation, which is subsequently used by fully connected network. Based on recent DeepSets architecture proposed by Zaheer et al. (NIPS 2017), we propose set aggregation network (SAN) as an alternative to pooling operation. In contrast to global pooling, SAN allows to embed a given set of features to a vector representation of arbitrary size. In consequence, by adjusting the size of embedding, SAN is capable of preserving the whole information from the input, which is also proven theoretically. In experiments, we demonstrate that replacing global pooling by SAN leads to the improvement of classification accuracy on various types of networks. Moreover, it is less prone to overfitting and can be used as a regularizer.
Abstract (translated by Google)
URL
http://arxiv.org/abs/1810.01868