Bayesian program learning

Bayesian Program Learning (BPL) proceeds in three steps:

1. In the first step, which is a generative model, BPL learns new concepts by building them compositionally from parts (refer to iii) of the A side in the diagram of the Model section), subparts (refer to ii) of the A side in the following diagram), and their spatial relations (refer to iv) of the A side in the following diagram). For example, it can sample new types of concepts (or, in this case, handwritten characters) from parts and subparts and combine them in new ways.
2. In the second step, the concepts sampled in the first step form another lower-level generative model to produce new examples as shown in the v) part of the A side.
3. The final step renders raw character level images. Hence, BPL is a generative model for generative models. The pseudocode for this generative model is shown on the B...