论文信息 - Replica-symmetry breaking in perceptrons

Replica-symmetry breaking in perceptrons

In the problem of optimization of pattern stabilization in perceptrons the replica-symmetric ansatz is known to be mathematically unstable for storage capacities greater than some . In this paper we demonstrate that for greater than the one-step replica-symmetry broken (RSB) solution is also unstable. We further show that in this region, full RSB is necessary for an exact solution. Direct evaluation of the two-step RSB solution yields a minimum storage error which is only slightly greater than the one-step RSB, which itself is greater than that given by the (unstable) replica-symmetric ansatz by a much larger amount.

D. Sherrington | W. Whyte

[1] D. Thouless,et al. Stability of the Sherrington-Kirkpatrick solution of a spin glass model , 1978 .

[2] Giorgio Parisi,et al. Infinite Number of Order Parameters for Spin-Glasses , 1979 .

[3] G. Parisi. The order parameter for spin glasses: a function on the interval 0-1 , 1980 .

[4] E. Gardner,et al. Optimal storage properties of neural network models , 1988 .

[5] Optimally adapted attractor neural networks in the presence of noise , 1990 .

[6] Stability of the one-step replica-symmetry-broken phase in neural networks , 1992 .

[7] Paul C. Bressloff,et al. Low firing-rates in a compartmental model neuron , 1993 .

[8] A. Engel,et al. Perceptrons above saturation , 1993 .

[9] R. Erichsen,et al. Optimal storage of a neural network model: a replica symmetry-breaking solution , 1993 .

[10] B. Derrida,et al. Replica symmetry instability in perceptron models , 1994 .