Probabilistic Programming in Scala

Probabilistic Programming in Scala

• 1.
  Probabilistic Programming inScala Guy Van den Broeck [email protected] Computer Science Department Katholieke Universiteit Leuven September 13, 2011
• 2.
  What is aPPL? Probabilistic graphical models (Bayesian networks) important in machine learning, statistics, robotics, vision, biology, neuroscience, artificial intelligence (AI) and cognitive science. Probabilistic Programming Languages unify general purpose programming with probabilistic modeling Examples Functional, extending Scheme (Church) or Scala (Figaro, FACTORIE, ScalaPPL) Logical, extending Prolog (ProbLog, PRISM, BLOG, Dyna) Extending C# (Infer.NET) Tasks Compute probabilities, most likely assignments given observations Learn parameters and programs Applications in natural language processing, computer vision, machine learning, bioinformatics, probabilistic planning, seismic monitoring, . . .
• 3.
  PPLs: The EasyCase Idea behind functional PPLs Any function (A, B, . . .) => R can also operate on probability distributions (Distr [A], Distr [B], . . .) => Distr [R] Starting Point Many library functions operate on Booleans: &&, ||, !, ==, != exists, forall Idea 1. Make a version of those functions that operates on distributions and returns a Distr[Boolean]. 2. Extend with probabilistic data structures. 3. Use them to model complex probability distributions.
• 4.
  PPLs: The EasyCase Idea behind functional PPLs Any function (A, B, . . .) => R can also operate on probability distributions (Distr [A], Distr [B], . . .) => Distr [R] Starting Point Many library functions operate on Booleans: &&, ||, !, ==, != exists, forall Idea 1. Make a version of those functions that operates on distributions and returns a Distr[Boolean]. 2. Extend with probabilistic data structures. 3. Use them to model complex probability distributions.
• 5.
  PPLs: The EasyCase Idea behind functional PPLs Any function (A, B, . . .) => R can also operate on probability distributions (Distr [A], Distr [B], . . .) => Distr [R] Starting Point Many library functions operate on Booleans: &&, ||, !, ==, != exists, forall Idea 1. Make a version of those functions that operates on distributions and returns a Distr[Boolean]. 2. Extend with probabilistic data structures. 3. Use them to model complex probability distributions.
• 6.
  PPLs: The EasyCase Idea behind functional PPLs Any function (A, B, . . .) => R can also operate on probability distributions (Distr [A], Distr [B], . . .) => Distr [R] Starting Point Many library functions operate on Booleans: &&, ||, !, ==, != exists, forall Idea 1. Make a version of those functions that operates on distributions and returns a Distr[Boolean]. 2. Extend with probabilistic data structures. 3. Use them to model complex probability distributions.
• 7.
  PPLs: The EasyCase Idea behind functional PPLs Any function (A, B, . . .) => R can also operate on probability distributions (Distr [A], Distr [B], . . .) => Distr [R] Starting Point Many library functions operate on Booleans: &&, ||, !, ==, != exists, forall Idea 1. Make a version of those functions that operates on distributions and returns a Distr[Boolean]. 2. Extend with probabilistic data structures. 3. Use them to model complex probability distributions.
• 8.
  Examples
• 9.
  Boolean formulae Random Variables are objects (each object independent) val a = Flip(0.3) val b = Flip(0.6) BooleanDistr has member functions that build Formula objects. val xor = a && !b || !a && b Run inference on BooleanDistr println("Probability = " + xor.probability()) Probability = 0.54
• 10.
  Boolean formulae Random Variables are objects (each object independent) val a = Flip(0.3) val b = Flip(0.6) BooleanDistr has member functions that build Formula objects. val xor = a && !b || !a && b Run inference on BooleanDistr println("Probability = " + xor.probability()) Probability = 0.54
• 11.
  Boolean formulae Random Variables are objects (each object independent) val a = Flip(0.3) val b = Flip(0.6) BooleanDistr has member functions that build Formula objects. val xor = a && !b || !a && b Run inference on BooleanDistr println("Probability = " + xor.probability()) Probability = 0.54
• 12.
  Boolean formulae Random Variables are objects (each object independent) val a = Flip(0.3) val b = Flip(0.6) BooleanDistr has member functions that build Formula objects. val xor = a && !b || !a && b Run inference on BooleanDistr println("Probability = " + xor.probability()) Probability = 0.54
• 13.
  2-state weather HMM abstract class Timestep { def rainy: BooleanDistr def umbrella = If(rainy, Flip(0.9), Flip(0.1)) } object StartState extends Timestep { val rainy = Flip(0.2) } class SuccessorState(predecessor: Timestep) extends Timestep { val rainy = If(predecessor.rainy, Flip(0.5), Flip(0.1)) } var timestep: Timestep = StartState for(i <- 1 until 2000) timestep = new SuccessorState(timestep) println("Probability = " + timestep.umbrella.probability())
• 14.
  2-state weather HMM abstract class Timestep { def rainy: BooleanDistr def umbrella = If(rainy, Flip(0.9), Flip(0.1)) } object StartState extends Timestep { val rainy = Flip(0.2) } class SuccessorState(predecessor: Timestep) extends Timestep { val rainy = If(predecessor.rainy, Flip(0.5), Flip(0.1)) } var timestep: Timestep = StartState for(i <- 1 until 2000) timestep = new SuccessorState(timestep) println("Probability = " + timestep.umbrella.probability())
• 15.
  2-state weather HMM abstract class Timestep { def rainy: BooleanDistr def umbrella = If(rainy, Flip(0.9), Flip(0.1)) } object StartState extends Timestep { val rainy = Flip(0.2) } class SuccessorState(predecessor: Timestep) extends Timestep { val rainy = If(predecessor.rainy, Flip(0.5), Flip(0.1)) } var timestep: Timestep = StartState for(i <- 1 until 2000) timestep = new SuccessorState(timestep) println("Probability = " + timestep.umbrella.probability())
• 16.
  Probabilistic Data Structures Probabilistic List: objects are in the list with a certain probability, given by a BooleanDistr trait ListDistr[T] extends Distribution[List[T]]{ def forall(f: T => BooleanDistr) : BooleanDistr def exists(f: T => BooleanDistr) : BooleanDistr } Replace all Boolean by BooleanDistr in member functions Many possibilities (Set,Tree,. . . )
• 17.
  Probabilistic Graphs Viral marketing Learning biological pathways Spread of influence in social networks class Person { val influencedFriends = new ListDistr[Person] def influences(target: Person): BooleanDistr = { if(target == this) True else friends.exists(_.influences(target)) } } val p1,p2,p3,p4,p5,p6 = new Person val influence1to2 = Flip(0.9) n1.influencedFriends += (influence1to2, p2) n2.influencedFriends += (influence1to2, p1) ... println("Probability = " + p1.influences(p4).probability())
• 18.
  Probabilistic Graphs Viral marketing Learning biological pathways Spread of influence in social networks class Person { val influencedFriends = new ListDistr[Person] def influences(target: Person): BooleanDistr = { if(target == this) True else friends.exists(_.influences(target)) } } val p1,p2,p3,p4,p5,p6 = new Person val influence1to2 = Flip(0.9) n1.influencedFriends += (influence1to2, p2) n2.influencedFriends += (influence1to2, p1) ... println("Probability = " + p1.influences(p4).probability())
• 19.
  Probabilistic Graphs Viral marketing Learning biological pathways Spread of influence in social networks class Person { val influencedFriends = new ListDistr[Person] def influences(target: Person): BooleanDistr = { if(target == this) True else friends.exists(_.influences(target)) } } val p1,p2,p3,p4,p5,p6 = new Person val influence1to2 = Flip(0.9) n1.influencedFriends += (influence1to2, p2) n2.influencedFriends += (influence1to2, p1) ... println("Probability = " + p1.influences(p4).probability())
• 20.
  Probabilistic Graphs Viral marketing Learning biological pathways Spread of influence in social networks class Person { val influencedFriends = new ListDistr[Person] def influences(target: Person): BooleanDistr = { if(target == this) True else friends.exists(_.influences(target)) } } val p1,p2,p3,p4,p5,p6 = new Person val influence1to2 = Flip(0.9) n1.influencedFriends += (influence1to2, p2) n2.influencedFriends += (influence1to2, p1) ... println("Probability = " + p1.influences(p4).probability())
• 21.
  Probabilistic Values Model any discrete distribution class ValDistr[T] extends Distribution[T]{ def ==(v: T): BooleanDistr def map[R](f: T => ValDistr[R]): ValDistr[R] } Apply any deterministic function to ValDistr arguments def Apply[A,R](f: (A) => R)(a: ValDistr[A]): ValDistr[R] def Apply[A,B,R](f: (A,B) => R) ... ... Example Two dice: probability that their sum is 8? val die1 = Uniform(1 to 6) val die2 = Uniform(1 to 6) val sum = Apply(_+_)(die1,die2) println("Probability = " + (sum == 8).probability())
• 22.
  Probabilistic Values Model any discrete distribution class ValDistr[T] extends Distribution[T]{ def ==(v: T): BooleanDistr def map[R](f: T => ValDistr[R]): ValDistr[R] } Apply any deterministic function to ValDistr arguments def Apply[A,R](f: (A) => R)(a: ValDistr[A]): ValDistr[R] def Apply[A,B,R](f: (A,B) => R) ... ... Example Two dice: probability that their sum is 8? val die1 = Uniform(1 to 6) val die2 = Uniform(1 to 6) val sum = Apply(_+_)(die1,die2) println("Probability = " + (sum == 8).probability())
• 23.
  Probabilistic Values Model any discrete distribution class ValDistr[T] extends Distribution[T]{ def ==(v: T): BooleanDistr def map[R](f: T => ValDistr[R]): ValDistr[R] } Apply any deterministic function to ValDistr arguments def Apply[A,R](f: (A) => R)(a: ValDistr[A]): ValDistr[R] def Apply[A,B,R](f: (A,B) => R) ... ... Example Two dice: probability that their sum is 8? val die1 = Uniform(1 to 6) val die2 = Uniform(1 to 6) val sum = Apply(_+_)(die1,die2) println("Probability = " + (sum == 8).probability())
• 24.
  3-state weather HMM object Sunny extends Weather object Foggy extends Weather object Rainy extends Weather abstract class Timestep { def weather: ValDistr[Weather] def umbrella = weather.map{ case Sunny => Flip(0.1) case Foggy => Flip(0.3) case Rainy => Flip(0.8) } } object StartState extends Timestep { val weather = ValDistr((0.3,Sunny), (0.3,Foggy), (0.4,Rainy)) } class SuccessorState(predecessor: Timestep) extends Timestep { val weather = predecessor.weather.map{ case Sunny => ValDistr((0.8,Sunny), (0.15,Foggy), (0.05,Rainy)) case Foggy => ValDistr((0.05,Sunny), (0.3,Foggy), (0.65,Rainy)) case Rainy => ValDistr((0.15,Sunny), (0.35,Foggy), (0.5,Rainy)) } }
• 25.
  Why Scala forProbabilistic Programming? Probabilistic programming as a library No separate compiler or VM mixin DSL Higher-order functions and generics Pass existing deterministic code to probabilistic model Memoization for efficient inference Object-oriented easy to model probabilistic databases Other advantages carry over to the probabilistic case
• 26.
  Thanks