Combined local search strategy for learning in networks of binary synapses

Haiping Huang; Haijun Zhou

doi:10.1209/0295-5075/96/58003

Home/Archive/arXiv

arXiv · 1105.1651

Combined local search strategy for learning in networks of binary synapses

Haiping Huang·Haijun Zhou

View original source ↗Read source document ↗Export metadata ↓

Abstract

Learning in networks of binary synapses is known to be an NP-complete problem. A combined stochastic local search strategy in the synaptic weight space is constructed to further improve the learning performance of a single random walker. We apply two correlated random walkers guided by their Hamming distance and associated energy costs (the number of unlearned patterns) to learn a same large set of patterns. Each walker first learns a small part of the whole pattern set (partially different for both walkers but with the same amount of patterns) and then both walkers explore their respective weight spaces cooperatively to find a solution to classify the whole pattern set correctly. The desired solutions locate at the common parts of weight spaces explored by these two walkers. The efficiency of this combined strategy is supported by our extensive numerical simulations and the typical Hamming distance as well as energy cost is estimated by an annealed computation.

Explore related subjects

cond-mat.dis-nn cond-mat.stat-mech cs.IT math.IT

Keep this discovery

Explore connections, maps & timelines

BibTeX RIS

Haiping Huang, Haijun Zhou. 2011-07-11. Combined local search strategy for learning in networks of binary synapses. https://doi.org/10.1209/0295-5075%2F96%2F58003

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Bias and Correlations in Quasiperiodicity: Impact on Localization in an Extended Aubry-André Model

Unlike in Anderson localization - where any amount of uncorrelated disorder localizes all eigenstates in one dimension - a one-dimensional system with quasiperiodic potential supports a richer range of localization behavior. This paper investigates the fundamental question of which potential characteristics govern localization properties. We characterize quasiperiodic potentials using two independent parameters, correlation and bias, and demonstrate that bias, in addition to correlation, critically influences localization. Through the study of an extended Aubry-André model, in which a tunable parameter allows systematic control over both the bias and correlation of the potential, we show that bias is key in determining the fraction of delocalized states. An increase in the potential strength generally enhances the tendency toward localization, while simultaneously strengthening the correlations in the quasiperiodic potential. This apparent counterintuitive behavior can be understood in terms of the bias parameter: increasing the potential strength reduces the bias, which in turn favors localization. For the family of Hamiltonians considered, we identify two critical bias thresholds: below the lower threshold, the entire spectrum is localized, whereas below the higher threshold, at most a fraction of the states can be delocalized, precluding delocalization of the entire spectrum. To further test our framework, we examine a quasiperiodic model with zero bias and find that, despite its correlated nature, all states become localized even at very weak potential strengths - recovering the Anderson-like scenario.

cond-mat.dis-nn↗

arXiv2026

Disorder-Tailored Delocalization

We derive disorder fields tailored by the details of a choice of a delocalized wave function. We first investigate the unidirectional Hatano-Nelson chain and its localization properties under $M$-base diagonal disorder with variable weights. The spectrum forms loops in the complex plane and the loop parameter is a good quantum number similar to a momentum. All eigenstates are subexponentially `localized', i.e. the logarithm of the absolute value of the wave function performs a random walk in space, and are characterized by a corresponding length scale $ξ_{sel}$ as shown in 1998 by Silvestrov for the general Hatano-Nelson chain. For $M=2$ real-valued binary disorder with equal weights the model was solved in Zeitschrift für Naturforschung A 81 421, yielding Cassini oval spectral loops and a diverging subexponential localization length for two eigenstates and for disorder weaker than a critical value set by the hopping strength. When the disorder field for any $M$ and arbitrary weights is confined to circles in the complex plane with radius equal to the hopping strength, the circle center will belong to the spectrum and to one of the spectral loops, and host a plane-wave-like eigenstate with diverging $ξ_{sel}$. We generalize to tailoring on-site disorder for a given eigenstate at a given energy, for any lattice dimension, and any hopping field (both ordered and disordered), for Hermitian and non-Hermitian systems. We exemplify by constructing a one-dimensional chain with Anderson-localized eigenstates hosting a completely delocalized one with random phases. The localization length diverges as $1/|E|^{2/3}$ upon approaching the delocalized state. Our method can be used for the systematic construction of disorder fields which host predefined eigenstates with arbitrary properties.

cond-mat.dis-nn↗

arXiv2026

Statistical levels and spatial modes of Fock-space heterogeneity in many-body localization crossovers

Near many-body localization crossovers, local memory fluctuates strongly among eigenstates and ensemble realizations, but the observed heterogeneity combines contributions from different statistical levels. We develop a statistical framework based on configuration-space distance distributions that uses a variance decomposition to separate fluctuations within eigenstates, between eigenstates of one sample, and between samples. Applying this framework to random-cosine and quasiperiodic-cosine Ising ensembles with the same one-site field marginal, we find that, at the exact-diagonalization sizes studied, the clearest difference occurs among disorder realizations or quasiperiodic phase samples, whereas within-eigenstate and within-sample eigenstate-to- eigenstate contributions remain broadly comparable. Spatial covariances show that random outer fluctuations have a much stronger uniform component, whereas quasiperiodic phase fluctuations are organized more strongly at finite wave number and partly cancel in the spatial average controlling the distance center. We find that, near the crossover in the random ensemble, the ensemble-averaged distance center is particularly sensitive to changes in the nominal field strength. Combined with sample-to-sample differences in the realized field amplitude, this mean response accounts for much of the sample-to-sample variation in the distance center. Analysis of half-chain entanglement further shows that its sample-to-sample fluctuations likewise reflect the combined effects of amplitude variations and its own mean response. An application to a fixed-magnetization spin chain demonstrates the framework in a constrained configuration space. Resolving both statistical level and spatial mode therefore provides a more complete picture of sample- dependent many-body memory and its configuration-space probability geometry.

cond-mat.dis-nn↗