Robustness of noise-induced synchronization Ngy 16 thng 4 nm 2008 - PowerPoint PPT Presentation

Robustness of noise-induced synchronization Ngày 16 tháng 4 năm 2008 () Noise-induced synchronization Ngày 16 tháng 4 năm 2008 1 / 25

Outline Problem statement 1 Modelling “noise”: Stochastic Differential Equations 2 The proof 3 Limitations of the analysis 4 () Noise-induced synchronization Ngày 16 tháng 4 năm 2008 2 / 25

Mainen & Sejnowski experiment [Mainen & Sejnowski, 1995] () Noise-induced synchronization Ngày 16 tháng 4 năm 2008 4 / 25

Synchronization interpretation () Noise-induced synchronization Ngày 16 tháng 4 năm 2008 5 / 25

Random walk and Wiener process Random walk (discrete-time): x t +∆ t = x t + ξ t ∆ t where ( ξ t ) t ∈ N are Gaussian and mutually independent If one is interested in very rapidly varying perturbations, ∆ t has to be very small Wiener process (or Brownian motion) (continuous-time): limit of the random walk when ∆ t → 0 15 10 5 0 −5 −10 0 2 4 6 8 10 () Noise-induced synchronization Ngày 16 tháng 4 năm 2008 7 / 25

Wiener process and “white noise” Problem: a Wiener process is not differentiable (why?), thus it is not the solution of any ordinary differential equation Define formally ξ t (“white noise”) = “derivative” of the Wiener process � t Formally: W ( t ) − W (0) = 0 ξ t dt or dW / dt = ξ t or dW = ξ t dt () Noise-induced synchronization Ngày 16 tháng 4 năm 2008 8 / 25

Stochastic Differential Equations We consider processes driven by “white noise” We would like to write (but it’s not correct, because ξ does not exist) x = f ( x ) + g ( x ) ξ ˙ In integral form, it may be more correct � t � t x ( t ) − x (0) = f ( x ) dt + g ( x ) dW 0 0 where the last term is a Stieltjes integral against W (which does exist) The integral form can also be written in differential form d x = f ( x ) dt + g ( x ) dW () Noise-induced synchronization Ngày 16 tháng 4 năm 2008 9 / 25

Definition of Itô and Stratonovich integrals For deterministic function α , the Stieltjes integral (which generalizes Riemann integrals) against α is defined as � T N − 1 � β ( t ) d α = lim β ( t i ) [ α ( t i +1 ) − α ( t i )] N →∞ 0 1 Thus one can define, by analogy � T N − 1 � β ( t ) dW = lim β ( t i ) [ W ( t i +1 ) − W ( t i )] N →∞ 0 1 which is the Itô integral But one can also define � T N − 1 � t i + t i +1 � � β ( t ) dW = lim β [ W ( t i +1 ) − W ( t i )] 2 N →∞ 0 1 which is the Stratonovich integral () Noise-induced synchronization Ngày 16 tháng 4 năm 2008 10 / 25

Independance properties The two above definitions lead to the same result in the deterministic case (probably, C 1 is required) But there are differences in the stochastic case: Since β ( t i ) (present) is independent of W ( t i +1 ) − W ( t i ) (future), one has, for Itô integrals E ( β ( t i )[ W ( t i +1 ) − W ( t i )]) = E ( β ( t i )) E ( W ( t i +1 ) − W ( t i )) = 0 leading to �� T � β ( t ) dW = 0 E 0 This explains Teramae claim “In the Itô formulation, [. . . ], the correlation between φ and ξ vanishes” () Noise-induced synchronization Ngày 16 tháng 4 năm 2008 11 / 25

Variable transformation The variable transformation (“changement de variable” in French) formula is also different for Itô and Stratonovich integrals Consider the function y ( x ) . In the deterministic case, one has, for instance dt = ∂ y dy = ∂ y dy ∂ x · dx ∂ x dx or dt The same rule is valid for Stratonovich integrals (Teramae’s “conventional variable transformation”): if dx = f ( x ) dt + g ( x ) dW then dy = ∂ y ∂ x ( f ( x ) dt + g ( x ) dW ) () Noise-induced synchronization Ngày 16 tháng 4 năm 2008 12 / 25

Itô’s formula for variable transformation Consider the Iô SDE dx = f ( x ) dt + g ( x ) dW Then for a function y ( x ) , one has (Itô’s formula) ∂ 2 y � ∂ y ∂ x f ( x ) + 1 � dt + ∂ y ∂ x 2 g ( x ) 2 dy = ∂ x g ( x ) dW 2 This will explain Teramae’s “the disappeared correlation is compensated by the new extra drift term Z ′ DZ ” () Noise-induced synchronization Ngày 16 tháng 4 năm 2008 13 / 25

Phase reduction Consider the system ˙ x = f ( x ) which has a limit-cycle. We would like to find a phase variable φ ( x ) such that: d φ dt = ω ω = constant x Example: a mobile travelling on a circle θ with constant velocity General case (using the chain rule): d φ dt = ∂φ dt = ∂φ ∂ x ( x ) · d x ∂ x ( x ) · f ( x ) = ω One then has to solve the above PDE to find φ () Noise-induced synchronization Ngày 16 tháng 4 năm 2008 15 / 25

Phase reduction(continued) Consider now a small perturbation ξ x = f ( x ) + ξ ˙ Then the equation on the phase becomes d φ dt = ∂φ ∂ x ( x ) · f ( x ) + ∂φ ∂ x ( x ) · ξ = ω + ∂φ ∂ x ( x ) · ξ This can be converted into a φ -only equation using some approximations d φ dt = ω + Z ( φ ) ξ This was equation (2) in [Teramae & Tanaka, 2004] () Noise-induced synchronization Ngày 16 tháng 4 năm 2008 16 / 25

Stratonovich to Itô switch Actually, the authors could have done everything in Itô! Let us compute the phase equation obtained previously but using now Itô’s formula (with D = 1 2 g 2 ) ∂ x f ( x ) + D ∂ 2 φ � ∂φ � dt + ∂φ d φ = ∂ x dW ∂ x 2 As above, let Z ( φ ) = ∂φ ∂ x . Then ∂ 2 φ ∂ x 2 = ∂ ∂ x Z ( φ ) = ∂ Z ∂φ ∂ x = Z ′ ( φ ) Z ( φ ) ∂φ Thus d φ = ( ω + Z ′ ( φ ) DZ ( φ )) dt + Z ( φ ) dW which is equation (3) (after formal division by dt ) () Noise-induced synchronization Ngày 16 tháng 4 năm 2008 17 / 25

Linearization Consider ˙ φ 1 = f ( φ 1 ) and ˙ φ 2 = f ( φ 2 ) Then (using the Taylor expansion assuming φ 1 − φ 2 very small) φ 1 − ˙ ˙ φ 2 = f ( φ 1 ) − f ( φ 2 ) = f ( φ 1 ) − ( f ( φ 1 ) + ( φ 2 − φ 1 ) f ′ ( φ 1 )) = f ′ ( φ 1 )( φ 1 − φ 2 ) This explains equation (4) if we set ψ = φ 1 − φ 2 () Noise-induced synchronization Ngày 16 tháng 4 năm 2008 18 / 25

Lyapunov exponent Consider two infinitesimally close trajectories. The Lyapunov exponent λ verifies (intuitively) � δφ ( t ) � ≃ e λ t � δφ 0 � If λ > 0 , then nearby trajectories diverge = instability If λ < 0 , then nearby the trajectories converge = stability (Remark: if a system is contracting, then λ < 0) () Noise-induced synchronization Ngày 16 tháng 4 năm 2008 19 / 25

Lyapunov exponent (continued) Let us manipulate the above expression: e λ t ≃ � δφ ( t ) � � � δφ ( t ) � � λ = 1 � � δφ ( t ) � � λ t = ln t ln � δφ 0 � � δφ 0 � � δφ 0 � Actually, the Lyapunov exponent is defined as (because we are interested in long-time behaviour) 1 � � δφ ( t ) � � λ = lim t ln � δφ 0 � t →∞ () Noise-induced synchronization Ngày 16 tháng 4 năm 2008 20 / 25

Ergodic hypothesis Consider a stochastic process x ( ω, t ) Any physicist knows that (ergodic hypothesis): ∀ ω, t � T 1 � x ( ω, t ′ ) dt ′ = x ( ω ′ , t ) d ω ′ = E ( x ( t )) lim T T →∞ 0 Ω Time average Ensemble average () Noise-induced synchronization Ngày 16 tháng 4 năm 2008 21 / 25

Ergodic hypothesis (continued) Remark now that y in equation (5) is defined as y = ln( ψ ) = ln( φ 1 − φ 2 ) = ln( δφ ) Remark that � T y = y ( T ) − y (0) = ln( δφ ( T )) − ln( δφ (0)) = ln δφ ( T ) ˙ δφ (0) 0 thus � T 1 y = 1 T →∞ ln δφ ( T ) lim ˙ lim δφ (0) = λ T T T →∞ 0 By the ergodic hypothesis, we then have λ = E (˙ y ) which explains the first line in equation (6). () Noise-induced synchronization Ngày 16 tháng 4 năm 2008 22 / 25

Probability density Let P ( φ, t ) denotes the time-dependent probability density of the random variable φ ∈ [0 , 2 π ] P is constant intuitively means that φ has equal probability of being anywhere in [0 , 2 π ] () Noise-induced synchronization Ngày 16 tháng 4 năm 2008 23 / 25

Some limitations A lot of unproven statements (ergodicity, uniform distribution of φ in steady state,. . . ). Perhaps those statements are evident for physicists! There is a mistake in the computation of the phase equation, as pointed out by [Yoshimura & Arai, 2007] (Thank you, Francis!). However, this mistake does not alter the result. The analysis is only valid when Z is continuously differentiable up to the second-order, which is not verified for e.g. resetting neuron models (Integrate and Fire, Izhikevich,...) () Noise-induced synchronization Ngày 16 tháng 4 năm 2008 25 / 25

Robustness of noise-induced synchronization Ngy 16 thng 4 nm 2008 - PowerPoint PPT Presentation

Robustness of noise-induced synchronization Ngy 16 thng 4 nm 2008 () Noise-induced synchronization Ngy 16 thng 4 nm 2008 1 / 25 Outline Problem statement 1 Modelling noise: Stochastic Differential Equations 2 The proof

Where Are We? Lecture 9 Robustness through Training 1 Robustness Explicit Handling of Noise

Too much noise to sleep: Noise-induced transition from sleep to awake-like state in a spiking

Robustness improvement of an SRAM cell against laser-induced fault injection A. Sarafianos a,b ,

Algorithms in Nature Network robustness Slides adapted from Carl Kingsford Network robustness

NOISE AT WORK AWARENESS SESSION FOR SUPERVISORS OBJECTIVES Understand what is noise

Reinforcing Adversarial Robustness using Model Confidence Induced by Adversarial Training Xi Wu

Assessment of noise induced hearing loss (NIHL) of mine workers in a bauxite mine using fuzzy

Loss factorization, weakly supervised learning and label noise robustness Giorgio Patrini,

Noise David Welch Plan Define Noise Overview Interesting bits of theory and our

Kramers type law for L L evy noise induced transitions Peter Imkeller Ilya Pavlyukevich

Noise-Induced Stop-and-Go Dynamics in Pedestrian Single-File Motion Andreas Schadschneider

Modularity, synchronization, and noise: a view from nonlinear contraction theory Quang-Cuong Pham

Robustness in GANs and in Black-box Optimization Stefanie Jegelka MIT CSAIL joint work with

Learning Hawkes Processes Under Synchronization Noise William Trouleau Jalal Etesami Negar

Stochastic Composite Optimization: Variance Reduction, Acceleration, and Robustness to Noise

Robustness? Robustness ? Robustness?

Processes on networks Robustness, resilience Random walks Diffusion, spreading

of Graph Embeddings Aleksandar Bojchevski Technical University of Munich, Germany Graph

Content Synchronization Content Synchronization March 2nd 2005 Jukka Honkola T-110.456

UCSD Robustness Summer School David Donoho 20190812 David Donoho UCSD Robustness Summer School

Robustness and Generalization Huan Xu The University of Texas at Austin Department of Electrical

Existence of Noise Induced Order, a computer aided proof. S. Galatolo Dip. Mat, Univ. Pisa CIRM

Robustness and SMC Adam Pechner Overview What is Robustness and why do we care? Different

Condition Synchronization 1 Synchronization Now that you have seen locks, is that all there is?