Sum Product Networks
What is a Sum Product Network? 1. It is a tractable probabilistic model (represented by a DAG) over a set of random variables with restrictions on what the structure can become. 2. Allows to compute marginal densities, MAP, likelihood, etc. 3. More importantly, computation of the marginals/partition function is O(n) (single bottom up pass through the network). 4. Additionally, we may marginalize over arbitrary subsets of variables as well. 5. Original paper was presented in 2011 by Poon & Domingos.
Diagram
Why Are They Useful ● Capable of intertwining with several other models (like VAE’s, CNNs, etc) ● Tractable Inference which is usually very difficult to guarantee ● Theoretical properties which justify certain methodology. ● Possibility for interpretation of structure (to some extent)
Possible Research Directions 1. Adding Features to SPFlow (which is an open source library built on top of tensorflow/pytorch) for SPNs 2. Adversarial Robustness in SPNs. 3. RSPMNs and Reinforcement Learning.
Contact Feel free to contact me after the lecture Email: pranav0212.s@gmail.com
Recommend
More recommend
