Annealed Neural Dynamics for Independent Component Analysis Jiann-Ming Wu Department of Applied Mathematics National Donghwa University Intelligent Numerical Computation 1 Outlines 1. 2. 3. 4. 5. 6. Blind source separation Independent component analysis Previous works, JadeICA, fastICA ICA using Potts models, PottsICA A comparison on the three methods Conclusions and future works Intelligent Numerical Computation 2 Blind Source Separation (BSS) 2 2 0 0 -2 0 2 20 40 60 80 100 120 140 160 180 200 0 -2 0 2 20 40 60 80 100 120 140 160 180 200 20 40 60 80 100 120 140 160 180 200 20 40 60 80 100 120 140 160 180 200 20 40 60 80 100 120 140 160 180 200 20 40 60 80 100 120 140 160 180 200 0 -2 0 2 20 40 60 80 100 120 140 160 180 200 0 -2 0 2 0 -2 0 2 20 40 60 80 100 120 140 160 180 200 0 -2 0 2 0 -2 0 2 20 40 60 80 100 120 140 160 180 200 0 -2 0 2 0 -2 -2 0 20 40 60 80 100 120 140 Sources 160 180 200 0 Unknown Mixing Structure Intelligent Numerical Computation Observations 3 BSS by PottsICA IEEE Trans. On Neural Networks(2001) PottsICA Observations 2 0 -2 0 2 20 40 60 80 100 120 140 160 180 200 20 40 60 80 100 120 140 160 180 200 0 20 40 60 80 100 120 140 160 180 200 -2 0 2 20 40 60 80 100 120 140 160 180 200 20 40 60 80 100 120 140 160 180 200 0 -2 0 2 0 -2 Recovered sources 2 0 0 -2 0 Intelligent Numerical Computation 4 Blind source separation: voice-music BSS for voice-music separation Intelligent Numerical Computation 5 BSS by ICA ICA Intelligent Numerical Computation 6 Blind Source separation : Fetal EKG ICA Intelligent Numerical Computation 7 Previous works FastICA:Helsinki University of Technology JadeICA:by JF Cardoso InformaxICA: Salk Institute The others… Intelligent Numerical Computation 8 The ICA problem Unknown mixing structure: Unkown statistical independent sources: S= Observations: Intelligent Numerical Computation 9 The goal of ICA Recover independent sources by such that The joint distribution is as close as possible to the product of the marginal distributions Intelligent Numerical Computation 10 Intelligent Numerical Computation 11 Intelligent Numerical Computation 12 The unmixing problem or Intelligent Numerical Computation 13 Intelligent Numerical Computation 14 Then Intelligent Numerical Computation 15 Partition the range of each output component … … Intelligent Numerical Computation 16 Non-overlapping projection the active state is where Intelligent Numerical Computation 17 Internal representations Intelligent Numerical Computation 18 Output component i Bin or state number … 1 2 3 4 k… K-1 K Membership vector =[0 0 0 0 … 1…0 0] the kth bit Intelligent Numerical Computation 19 A mathematical programming for non-overlapping projection Intelligent Numerical Computation 20 Statistical mechanism Intelligent Numerical Computation 21 Normalized histogram Intelligent Numerical Computation 22 Discrete marginal entropy Intelligent Numerical Computation 23 A mathematical programming for ICA Intelligent Numerical Computation 24 Energy function for ICA To minimize L is to solve a mixed integer and linear programming Intelligent Numerical Computation 25 Annealed neural dynamics Boltzmann distribution Use mean field equations to find the mean configuration at each Intelligent Numerical Computation 26 Derivation of mean field equations Free energy by Intelligent Numerical Computation 27 Mean field equations Intelligent Numerical Computation 28 A hybrid of mean field annealing and gradient descent method MFE Intelligent Numerical Computation 29 A hybrid of mean field annealing and gradient descent method Gradient descent method Intelligent Numerical Computation 30 The PottsICA algorithm Intelligent Numerical Computation 31 Numerical simulations Artificial problem I Independent sources Intelligent Numerical Computation 32 Numerical simulations The normalized histogram of the five sources Intelligent Numerical Computation 33 Numerical simulations The mixed signals of the five sources using a randomly generated mixing matrix Intelligent Numerical Computation 34 Performance evaluations by Amari Intelligent Numerical Computation 35 Performance evaluations Intelligent Numerical Computation 36 Recovered signals by PottsICA Intelligent Numerical Computation 37 Numerical simulations The normalized histograms of the eight sources Gaussian Sub-Gaussian Super-Gaussian Intelligent Numerical Computation 38 Performance evaluations Intelligent Numerical Computation 39 Performance evaluations Intelligent Numerical Computation 40 Conclusions and Future works 1. Apply Potts encoding to ICA 2. PottsICA is better than JadeICA and FastICA 3. PottsICA is of collective decisions, PDP 4. PottsICA is potential for real world applications Intelligent Numerical Computation 41
© Copyright 2026 Paperzz