Deep Generalized Convolutional Sum-Product Networks for Probabilistic Image Representations

Sum-Product Networks (SPNs) are hierarchical, probabilistic graphical models capable of fast and exact inference that can be trained directly from high-dimensional, noisy data. Traditionally, SPNs struggle with capturing relationships in complex spatial data such as images. To this end, we introduce Deep Generalized Convolutional Sum-Product Networks (DGC-SPNs), which encode spatial features through products and sums with scopes corresponding to local receptive fields. As opposed to existing convolutional SPNs, DGC-SPNs allow for overlapping convolution patches through a novel parameterization of dilation and strides, resulting in significantly improved feature coverage and feature resolution. DGC-SPNs substantially outperform other convolutional and non-convolutional SPN approaches across several visual datasets and for both generative and discriminative tasks, including image completion and image classification. In addition, we demonstrate a modificiation to hard EM learning that further improves the generative performance of DGC-SPNs. While fully probabilistic and versatile, our model is scalable and straightforward to apply in practical applications in place of traditional deep models. Our implementation is tensorized, employs efficient GPU-accelerated optimization techniques, and is available as part of an open-source library based on TensorFlow.