Capacity of Hopfield Memory • Depends on the patterns imprinted • If orthogonal, pmax = n – but every state is stable fi trivial basins • So pmax < n • Let load parameter a = p / n 11/5/03 equations 1 Single Bit Stability Analysis • For simplicity, suppose xk are random • Then xk ⋅ xm are sums of n random ±1 ß ß ß ß binomial distribution ≈ Gaussian in range –n, …, +n with mean m = 0 and variance s2 = n • Probability sum > t: È ˘ Ê ˆ t 1 ˜˙ 2 Í1- erf Á Ë 2n ¯˚ Î [See “Review of Gaussian (Normal) Distributions” on course website] 11/5/03 2 † Approximation of Probability Let crosstalk Cim = 1 n k k m x x ⋅ x  i( ) k≠m We want Pr{Cim > 1} = Pr{nCim > n} p † † † † † n Note : nCim =   x ik x kj x mj k=1 j=1 k≠m A sum of n( p -1) ª np random ± 1 Variance s 2 = np 11/5/03 3 Probability of Bit Instability È ˘ Ê ˆ n Pr{nCim > n} = 12 Í1- erf Á ˜˙ ÍÎ Ë 2np ¯˙˚ = 1 2 [1- erf ( n 2p )] † 11/5/03 (fig. from Hertz & al. Intr. Theory Neur. Comp.) 4 Tabulated Probability of Single-Bit Instability a Perror 11/5/03 0.1% 0.105 0.36% 0.138 1% 0.185 5% 0.37 10% 0.61 (table from Hertz & al. Intr. Theory Neur. Comp.) 5 Spurious Attractors • Mixture states: – – – – sums or differences of odd numbers of retrieval states number increases combinatorially with p shallower, smaller basins basins of mixtures swamp basins of retrieval states fi overload – useful as combinatorial generalizations? – self-coupling generates spurious attractors • Spin-glass states: – not correlated with any finite number of imprinted patterns – occur beyond overload because weights effectively random 11/5/03 6 Basins of Mixture States x k1 x x † † 11/5/03 k3 mix † k2 x x imix = sgn( x ik1 + x ik2 + x ik3 ) 7 Fraction of Unstable Imprints (n = 100) 11/5/03 (fig from Bar-Yam) 8 Number of Stable Imprints (n = 100) 11/5/03 (fig from Bar-Yam) 9 Number of Imprints with Basins of Indicated Size (n = 100) 11/5/03 (fig from Bar-Yam) 10 Summary of Capacity Results • Absolute limit: pmax < acn = 0.138 n • If a small number of errors in each pattern permitted: pmax µ n • If all or most patterns must be recalled perfectly: pmax µ n / log n • Recall: all this analysis is based on random patterns • Unrealistic, but sometimes can be arranged 11/5/03 11 Stochastic Neural Networks (in particular, the stochastic Hopfield network) 11/5/03 12 Trapping in Local Minimum 11/5/03 13 Escape from Local Minimum 11/5/03 14 Escape from Local Minimum 11/5/03 15 Motivation • 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 11/5/03 16 The Stochastic Neuron Deterministic neuron : s¢i = sgn( hi ) Pr{s¢i = +1} = Q( hi ) Pr{s¢i = -1} = 1- Q( hi ) Stochastic neuron : Pr{s¢i = +1} = s ( hi ) † † † s(h) h Pr{s¢i = -1} = 1- s ( hi ) 1 Logistic sigmoid : s (h ) = 1+ exp(-2 h T ) 11/5/03 17 Properties of Logistic Sigmoid 1 s (h) = 1+ e-2h T • As h Æ +•, s(h) Æ 1 • As h Æ –•, s(h)†Æ 0 • s(0) = 1/2 11/5/03 18 Logistic Sigmoid With Varying T 11/5/03 19 Logistic Sigmoid T = 0.5 Slope at origin = 1 / 2T 11/5/03 20 Logistic Sigmoid T = 0.01 11/5/03 21 Logistic Sigmoid T = 0.1 11/5/03 22 Logistic Sigmoid T=1 11/5/03 23 Logistic Sigmoid T = 10 11/5/03 24 Logistic Sigmoid T = 100 11/5/03 25