Exploring Sum-Product Networks for Image Classification. Applications

Sum-Product Networks (SPN) represent a significant and novel development within the probabilistic graphical models domain that unites the representational capabilities of deep learning with effective probabilistic inference. Our focus is on presenting these unique structural properties of SPNs along with their theoretical foundations and putting them into application on real-life diverse datasets. Despite its promising prospects, only a few existing applications are using this technology. To illustrate how this method can be used in practice, we have applied the model to four different datasets: the popular MNIST dataset, the Fashion MNIST dataset, the Fast, Furious and Insured dataset, and the Garbage Classification dataset, respectively. In the first two datasets, we used SPN s to classify hand-written digits and fashion products respectively displaying their effectiveness within a standard image classification task. The models' architecture incorporates SPN-specific layers together with conventional convolutional layers thereby achieving remarkable performance measures. The results obtained for the last two datasets outperformed the results obtained by the classical deep learning methods. All these show that SPNs form a strong basis for theoretical exploration and practical application to probabilistic modeling along with deep learning. This paper adds to the collective understanding of SPN applications and, at the same time, highlights possible further areas of study within the field.