Probabilistic Circuit Units

diff computational graph

• encode a fn over X

Units in Computational Graph

Types

• input unit (any fn) : distribution
• product unit : factorization
• sum unit : mixture
• integral unit
  • continuous heirarchial mixture models
  • stack K gaussian components
  • output is a function after solving an analytical integration
• can be stacked
• outputs a number
• evaluated bottom up - joint probability
• non linearities at input and product
• probabilistic
  • when it encodes a non negative fn
  • input units encode distributions (like Normal Distribution), sum weights are positive
• for tractability
  • sum/products must be smooth/decomposable
  • can compute marginal probability (exact)