Distinguishing Spurious and Nominal Attractors Applying Unlearning to an Asymmetric Neural Network

Horas, J. A.; Bea, Edgar A.

doi:10.1142/s0129065702001047

Cited by 2 publications

(1 citation statement)

References 11 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…There have been many studies on unlearning in neural networks. [6][7][8][9][10] These papers treat unlearning the final states starting from noisy inputs, or unlearning by parallel dynamics, or unlearning by using asymmetric synaptic weights, and so forth. In particular, in Refs.…”

Section: Introductionmentioning

confidence: 99%

Unlearning of Mixed States in the Hopfield Model —Finite Loading Case—

et al. 2015

View full text Add to dashboard Cite

We study the unlearning of mixed states in the Hopfield model for the finite loading case, that is, α = p N 1, where N and p are the numbers of neurons and embedded patterns, respectively. In the general situation that any number of mixed states that exist in the model is unlearned, we derive the saddle point equations (SPEs) and evolution equations for overlaps by introducing sublattices. We postulate a condition that the solutions are stable in equilibrium, and prove that the static and dynamic stabilities are the same. We also prove that the stable state of the Hopfield model continuously changes and is statically and dynamically stable for sufficiently small unlearning coefficients. For p = 3, we perform detailed theoretical and numerical calculations. In the case that a single mixed state is unlearned, we determine phase boundaries using the Hessian matrix and by numerically integrating evolution equations. We performed Markov chain Monte Carlo simulations and find that the simulation results agree with the theoretical ones reasonably well. For general p, when all of the mixed states are unlearned with an equal unlearning coefficient η, we derive the formulae for critical unlearning coefficients at the temperature T = 0 below which embedded patterns and mixed states exist and are stable as solutions of the SPEs. We found that there is an unlearning region of (T, η) in which all patterns are retained and all mixed states are deleted, although tuning the parameters in this region is more difficult as p increases since the region shrinks. We numerically confirmed the theoretical results of the p dependences of the critical unlearning coefficients.

show abstract