Abstract
Finding a high probability mode of a probability distribution is an important step in many machine-learning tasks such as structured prediction. In fact, this was exactly the inference used in the seminal paper by Poon and Domingos to show the effectiveness of Sum-Product Networks. Very often, one is also interested in finding many relevant modes, for example, when one is interested in analizying the robustness of the particular mode to perturbations in the model or when one seeks producing a diverse set of representative solutions to the task. There has been very few work and very limited progress in finding and analyzing modes of non-determinstic Probabilistic Circuits. In this talk, I will discuss the motivation, challenges, and tractability of finding modes in Probabilistic Circuits, with focus on Gaussian Sum-Product Networks.