expressive power of evolving neural networks working on
play

Expressive Power of Evolving Neural Networks Working on Infinite - PowerPoint PPT Presentation

Introduction RNNs Muller TMs Topology Det. -NNs Nondet. -RNNs Conclusion Expressive Power of Evolving Neural Networks Working on Infinite Input Streams Olivier Finkel 1 emie Cabessa 2 Joint work with J er 1 CNRS and University


  1. Introduction RNNs Muller TMs Topology Det. ω -NNs Nondet. ω -RNNs Conclusion Expressive Power of Evolving Neural Networks Working on Infinite Input Streams Olivier Finkel 1 emie Cabessa 2 Joint work with J´ er´ 1 CNRS and University Paris 7 2 Laboratoire d’´ Economie Math´ ematique, Universit´ e Paris 2 FCT 2017, Bordeaux 12 Septembre 2017 Expressive Power of Evolving Neural Networks J´ er´ emie Cabessa & Olivier Finkel

  2. Introduction RNNs Muller TMs Topology Det. ω -NNs Nondet. ω -RNNs Conclusion Introduction ◮ The computational capabilities of recurrent neural networks have mainly been studied in the context of classical computa- tion: McCulloch & Pitts (1943), Turing (1948), Kleene (1956), von Neumann (1958), Minsky (1967), Papert (1969),. . . , Siegel- mann & Sontag (1994-1995),. . . ◮ We provide a characterization of the computational power of recurrent neural networks in terms of their attractor dynamics, i.e., in the context of infinite input stream computation. Expressive Power of Evolving Neural Networks J´ er´ emie Cabessa & Olivier Finkel

  3. Introduction RNNs Muller TMs Topology Det. ω -NNs Nondet. ω -RNNs Conclusion Introduction ◮ The computational capabilities of recurrent neural networks have mainly been studied in the context of classical computa- tion: McCulloch & Pitts (1943), Turing (1948), Kleene (1956), von Neumann (1958), Minsky (1967), Papert (1969),. . . , Siegel- mann & Sontag (1994-1995),. . . ◮ We provide a characterization of the computational power of recurrent neural networks in terms of their attractor dynamics, i.e., in the context of infinite input stream computation. Expressive Power of Evolving Neural Networks J´ er´ emie Cabessa & Olivier Finkel

  4. Introduction RNNs Muller TMs Topology Det. ω -NNs Nondet. ω -RNNs Conclusion Recurrent Neural Network w9 w2 w13 w3 w15 w7 w11 w16 w4 w5 w12 w10 w14 w1 w6 w8 Expressive Power of Evolving Neural Networks J´ er´ emie Cabessa & Olivier Finkel

  5. Introduction RNNs Muller TMs Topology Det. ω -NNs Nondet. ω -RNNs Conclusion Boolean Neural Networks a i 1 a i 2 a iN neuron x i b i 1 1 θ 0 1 b iM c i   N M � � a ij · x j ( t ) + b ij · u j ( t ) + c i x i ( t + 1) = θ   j =1 j =1 Expressive Power of Evolving Neural Networks J´ er´ emie Cabessa & Olivier Finkel

  6. Introduction RNNs Muller TMs Topology Det. ω -NNs Nondet. ω -RNNs Conclusion Sigmoidal Neural Networks a i 1 a i 2 a iN neuron x i b i 1 1 σ 0 1 b iM c i   N M � � a ij · x j ( t ) + b ij · u j ( t ) + c i x i ( t + 1) = σ   j =1 j =1 Expressive Power of Evolving Neural Networks J´ er´ emie Cabessa & Olivier Finkel

  7. Introduction RNNs Muller TMs Topology Det. ω -NNs Nondet. ω -RNNs Conclusion Sigmoidal Neural Networks a i 1 ( t ) a i 2 ( t ) a iN ( t ) neuron x i b i 1 ( t ) 1 σ 0 1 b iM ( t ) c i ( t )   N M � � a ij ( t ) · x j ( t ) + b ij ( t ) · u j ( t ) + c i ( t ) x i ( t + 1) = σ   j =1 j =1 Expressive Power of Evolving Neural Networks J´ er´ emie Cabessa & Olivier Finkel

  8. Introduction RNNs Muller TMs Topology Det. ω -NNs Nondet. ω -RNNs Conclusion Neural Networks We consider three models of NNs: 1. Boolean rational NNs: B-NN[ Q ]s 2. Sigmoidal static rational NNs: St-NN[ Q ]s 3. Sigmoidal bi-valued evolving rational NNs: Ev 2 -NN[ Q ]s Expressive Power of Evolving Neural Networks J´ er´ emie Cabessa & Olivier Finkel

  9. Introduction RNNs Muller TMs Topology Det. ω -NNs Nondet. ω -RNNs Conclusion Neural Networks We consider three models of NNs: 1. Boolean rational NNs: B-NN[ Q ]s 2. Sigmoidal static rational NNs: St-NN[ Q ]s 3. Sigmoidal bi-valued evolving rational NNs: Ev 2 -NN[ Q ]s a i 1 a i 2 a iN neuron x i b i 1 1 θ 0 1 b iM c i Expressive Power of Evolving Neural Networks J´ er´ emie Cabessa & Olivier Finkel

  10. Introduction RNNs Muller TMs Topology Det. ω -NNs Nondet. ω -RNNs Conclusion Neural Networks We consider three models of NNs: 1. Boolean rational NNs: B-NN[ Q ]s 2. Sigmoidal static rational NNs: St-NN[ Q ]s 3. Sigmoidal bi-valued evolving rational NNs: Ev 2 -NN[ Q ]s a i 1 a i 2 a iN neuron x i b i 1 1 σ 0 1 b iM c i Expressive Power of Evolving Neural Networks J´ er´ emie Cabessa & Olivier Finkel

  11. Introduction RNNs Muller TMs Topology Det. ω -NNs Nondet. ω -RNNs Conclusion Neural Networks We consider three models of NNs: 1. Boolean rational NNs: B-NN[ Q ]s 2. Sigmoidal static rational NNs: St-NN[ Q ]s 3. Sigmoidal bi-valued evolving rational NNs: Ev 2 -NN[ Q ]s a i 1 ( t ) a i 2 ( t ) a iN ( t ) neuron x i b i 1 ( t ) 1 σ 0 1 b iM ( t ) c i ( t ) Expressive Power of Evolving Neural Networks J´ er´ emie Cabessa & Olivier Finkel

  12. Introduction RNNs Muller TMs Topology Det. ω -NNs Nondet. ω -RNNs Conclusion Results (Classical Computation) Boolean Static Bi-valued Evolving FSA TM TM/poly(A) REG P P/poly Kleene 56 Siegelmann & Cabessa & Minsky 67 Sontag 95 Siegelmann 11,14 Expressive Power of Evolving Neural Networks J´ er´ emie Cabessa & Olivier Finkel

  13. Introduction RNNs Muller TMs Topology Det. ω -NNs Nondet. ω -RNNs Conclusion Muller Turing Machine A Muller Turing machine consists of a classical TM with Muller acceptance condition. Muller table T : collection of accepting sets of states. ◮ The ω -word u is accepted by M if there is an infinite run ρ u of the machine M on u such that inf( ρ u ) ∈ T ◮ The ω -language accepted by M is the set of ω -words accepted by M . Expressive Power of Evolving Neural Networks J´ er´ emie Cabessa & Olivier Finkel

  14. Introduction RNNs Muller TMs Topology Det. ω -NNs Nondet. ω -RNNs Conclusion Complexity of ω -languages The question naturally arises of the complexity of ω -languages accepted by various kinds of automata. A way to study the complexity of ω -languages is to consider their topological complexity. Expressive Power of Evolving Neural Networks J´ er´ emie Cabessa & Olivier Finkel

  15. Introduction RNNs Muller TMs Topology Det. ω -NNs Nondet. ω -RNNs Conclusion Topology on Σ ω The natural prefix metric on the set Σ ω of ω -words over Σ is defined as follows: For u, v ∈ Σ ω and u � = v let δ ( u, v ) = 2 − n where n is the least integer such that: the ( n + 1) st letter of u is different from the ( n + 1) st letter of v . This metric induces on Σ ω the usual Cantor topology for which : ◮ open subsets of Σ ω are in the form W. Σ ω , where W ⊆ Σ ⋆ . ◮ closed subsets of Σ ω are complements of open subsets of Σ ω . Expressive Power of Evolving Neural Networks J´ er´ emie Cabessa & Olivier Finkel

  16. Introduction RNNs Muller TMs Topology Det. ω -NNs Nondet. ω -RNNs Conclusion Borel Hierarchy Σ 0 1 is the class of open subsets of Σ ω , Π 0 1 is the class of closed subsets of Σ ω , For any countable ordinal α ≥ 2 : α is the class of countable unions of subsets of Σ ω in � Σ 0 γ<α Π 0 γ . Π 0 α is the class of complements of Σ 0 α -sets ∆ 0 α = Π 0 α ∩ Σ 0 α . A set X ⊆ Σ ω is a Borel set iff it is in � α<ω 1 Σ 0 α<ω 1 Π 0 α = � α where ω 1 is the first uncountable ordinal. Expressive Power of Evolving Neural Networks J´ er´ emie Cabessa & Olivier Finkel

  17. Introduction RNNs Muller TMs Topology Det. ω -NNs Nondet. ω -RNNs Conclusion Borel Hierarchy Below an arrow → represents a strict inclusion between Borel classes. Π 0 Π 0 Π 0 1 α α +1 ր ց ր ր ց ր ∆ 0 ∆ 0 ∆ 0 ∆ 0 · · · · · · · · · 1 2 α α +1 ց ր ց ց ր ց Σ 0 Σ 0 Σ 0 1 α α +1 Expressive Power of Evolving Neural Networks J´ er´ emie Cabessa & Olivier Finkel

  18. Introduction RNNs Muller TMs Topology Det. ω -NNs Nondet. ω -RNNs Conclusion Beyond the Borel Hierarchy There are some subsets of Σ ω which are not Borel. Beyond the Borel hierarchy is the projective hierarchy. The class of Borel subsets of Σ ω is strictly included in the class Σ 1 1 of analytic sets which are obtained by projection of Borel sets. A set E ⊆ Σ ω is in the class Σ 1 1 iff : ∃ F ⊆ (Σ × { 0 , 1 } ) ω such that F is Π 0 2 and E is the projection of F onto Σ ω A set E ⊆ Σ ω is in the class Π 1 1 iff Σ ω − E is in Σ 1 1 . Suslin’s Theorem states that : Borel sets = ∆ 1 1 = Σ 1 1 ∩ Π 1 1 Expressive Power of Evolving Neural Networks J´ er´ emie Cabessa & Olivier Finkel

  19. Introduction RNNs Muller TMs Topology Det. ω -NNs Nondet. ω -RNNs Conclusion Deterministic ω -NNs We consider RNNs with Boolean input and output cells, sigmoidal internal cells, and working on infinite input streams. Boolean Sigmoid Boolean · · · input internal output · · · cells cells cells Infinite Boolean · · · input stream Infinite Boolean output stream Attractor (periodic) The attractors are assumed to be classified into two possible kinds: accepting or rejecting . Expressive Power of Evolving Neural Networks J´ er´ emie Cabessa & Olivier Finkel

Recommend


More recommend