Probabilistic Neural-symbolic Models for Interpretable ...

Given an image (i), we build a model for p(x;z;aji) (see Figure1); where the model factorizes as: p(x;z;aji) = p( z )x ja ;i . The implied generative process is as follows. First we sample a program z, which generates questions x. Further, given a program z and an image i we generate answers a. Note that based on the symbolic pro- ................
................