Experiments on PMAP learning of RBM

[StannisZhou]

Setup

Learning RBMs on MNIST digits from Sec. 5.5 of PMAP paper

Metrics

1. Sampling quality (Fig. 3(a) of PMAP paper)
2. Visualization of samples (Fig. 3(b) of PMAP paper)
3. Speed of inference

Concrete TODOs

• Implement perturb-and-max-product in PGMax @StannisZhou
• Implement perturb-and-map using PGMPy (candidate methods are MPLP and junction tree based BP) @NishanthJKumar
• Implement perturb-and-map using pomegranate (using the predict method mentioned in tutorial which seems to be doing MAP inference) @NishanthJKumar
• [ ] Integrate all perturb-and-map implementations and learn and sample from the models @StannisZhou