Motivation. Lecture 23. The Stochastic Neuron ( ) Properties of Logistic Sigmoid. Logistic Sigmoid With Varying T. Logistic Sigmoid T = 0.

Transcription

1 Motivation Lecture 23 Idea: with low probability, go against the local field move up the energy surface make the wrong microdecision Potential value for optimization: escape from local optima Potential value for associative memory: escape from spurious states because they have higher energy than imprinted states /2/07 /2/07 2 The Stochastic Neuron Deterministic neuron : s " i = sgn h i Pr { s " i = +} = # h i Pr { s " i = $} =$ # h i Stochastic neuron : Pr { s " i = +} = #( h i ) Pr { s " i = $} =$# h i Logistic sigmoid : " h = + exp #2 h T σ(h) h Properties of As h +, σ(h) As h, σ(h) 0 σ(0) = /2 = " h + e #2h T /2/07 3 /2/07 4 With Varying T T = 0.5 T varying from 0.05 to (/T = β = 0,, 2,, 20) Slope at origin = / 2T /2/07 5 /2/07 6

2 T = 0.0 T = 0. /2/07 7 /2/07 8 T = T = 0 /2/07 9 /2/07 0 T = 00 Pseudo-Temperature Temperature = measure of thermal energy (heat) Thermal energy = vibrational energy of molecules A source of random motion Pseudo-temperature = a measure of nondirected (random) change Logistic sigmoid gives same equilibrium probabilities as Boltzmann-Gibbs distribution /2/07 /2/07 2 2

3 Transition Probability Recall, change in energy "E = #"s k h k = 2s k h k Pr { s " k = ±s k = m} = #( ±h k ) = #( $s k h k ) Pr{ s k " #s k } = + exp 2s k h k T = + exp $E T /2/07 3 Stability Are stochastic Hopfield nets stable? Thermal noise prevents absolute stability But with symmetric weights: average values s i become time - invariant /2/07 4 Does Thermal Noise Improve Memory Performance? Experiments by Bar-Yam (pp ): n = 00 p = 8 Random initial state To allow convergence, after 20 cycles set T = 0 How often does it converge to an imprinted pattern? /2/07 5 Probability of Random State Converging on Imprinted State (n=00, p=8) T = / β /2/07 (fig. from Bar-Yam) 6 Probability of Random State Converging on Imprinted State (n=00, p=8) Analysis of Stochastic Hopfield Network Complete analysis by Daniel J. Amit & colleagues in mid-80s See D. J. Amit, Modeling Brain Function: The World of Attractor Neural Networks, Cambridge Univ. Press, 989. The analysis is beyond the scope of this course /2/07 (fig. from Bar-Yam) 7 /2/07 8 3

4 Phase Diagram (D) all states melt Conceptual Diagrams of Energy Landscape (C) spin-glass states (A) imprinted = minima (B) imprinted, but s.g. = min. /2/07 (fig. from Domany & al. 99) 9 /2/07 (fig. from Hertz & al. Intr. Theory Neur. Comp.) 20 Phase Diagram Detail Simulated Annealing (Kirkpatrick, Gelatt & Vecchi, 983) /2/07 (fig. from Domany & al. 99) 2 /2/07 22 Dilemma In the early stages of search, we want a high temperature, so that we will explore the space and find the basins of the global minimum In the later stages we want a low temperature, so that we will relax into the global minimum and not wander away from it Solution: decrease the temperature gradually during search /2/07 23 Quenching vs. Annealing Quenching: rapid cooling of a hot material may result in defects & brittleness local order but global disorder locally low-energy, globally frustrated Annealing: slow cooling (or alternate heating & cooling) reaches equilibrium at each temperature allows global order to emerge achieves global low-energy state /2/

5 Multiple Domains Moving Domain Boundaries global incoherence local coherence /2/07 25 /2/07 26 Effect of Moderate Temperature Effect of High Temperature /2/07 (fig. from Anderson Intr. Neur. Comp.) 27 /2/07 (fig. from Anderson Intr. Neur. Comp.) 28 Effect of Low Temperature Annealing Schedule Controlled decrease of temperature Should be sufficiently slow to allow equilibrium to be reached at each temperature With sufficiently slow annealing, the global minimum will be found with probability Design of schedules is a topic of research /2/07 (fig. from Anderson Intr. Neur. Comp.) 29 /2/

6 Typical Practical Annealing Schedule Initial temperature T 0 sufficiently high so all transitions allowed Exponential cooling: T k+ = αt k typical 0.8 < α < 0.99 at least 0 accepted transitions at each temp. Final temperature: three successive temperatures without required number of accepted transitions Summary Non-directed change (random motion) permits escape from local optima and spurious states Pseudo-temperature can be controlled to adjust relative degree of exploration and exploitation /2/07 3 /2/07 32 Additional Bibliography. Anderson, J.A. An Introduction to Neural Networks, MIT, Arbib, M. (ed.) Handbook of Brain Theory & Neural Networks, MIT, Hertz, J., Krogh, A., & Palmer, R. G. Introduction to the Theory of Neural Computation, Addison-Wesley, 99. V. Genetics & Evolution /2/07 33 /2/07 34 The Second Law of Thermodynamics The Second Law and Open Systems closed system entropy H H open system energy concentration H waste /2/07 35 /2/

7 Nonequilibrium Thermodynamics Classical thermodynamics limited to systems in equilibrium Extended by thermodynamics of transport processes i.e. accounting for entropy changes when matter/energy transported into or out of an open system Flow of matter/energy can maintain a dissipative system far from equilibrium for long periods Hence, nonequilibrium thermodynamics An Energy Flow Can Create Structure /2/07 37 /2/07 (photo from Camazine & al. Self-Org. Bio. Sys.) 38 Bénard Convection Cells Persistent Nonequilibrium Systems If flow creates system so structured to maintain flow then positive feedback causes nonequilibrium system to persist indefinitely but not forever (2 nd law) Systems we tend to see are those most successful at maintaining nonequil. state Applies to species as well as organisms /2/07 (photo from Camazine & al. Self-Org. Bio. Sys.) 39 /2/07 40 Order Through Fluctuations Stabilization at Successively Higher Levels of Organization Fluctuations (esp. when system forced out of ordinary operating range) test boundaries & nonlinear effects May lead to stabilization of new structures /2/07 4 fig. < Hart & Gregor, Inf. Sys. Found. /2/07 42 fig. < Hart & Gregor, Inf. Sys. Found. 7

8 Order for Free Relatively simple sets of rules or equations can generate rich structures & behaviors Small changes can lead to qualitatively different structures & behaviors A diverse resource for selection A basis for later fine tuning (microevolution) See Kaufmann (At Home in the Universe, etc.) and Wolfram (A New Kind of Science) Evolution in Broad Sense Evolution in the broadest terms: blind variation selective retention Has been applied to nonbiological evolution evolutionary epistemology creativity memes /2/07 43 /2/07 44 Evolution atoms & molecules replicating molecules living things prebiotic evolution biotic evolution /2/07 45 /2/07 (from NECSI) 46 8