Metastability in the generalized Hopfield model with finitely many patterns

نویسنده

  • Mykhaylo Shkolnikov
چکیده

This paper continues the study of metastable behaviour in disordered mean field models initiated in [2], [3]. We consider the generalized Hopfield model with finitely many independent patterns ξ1, . . . , ξP where the patterns have i.i.d. components and the components of patterns ξ1, . . . ξp have absolutely continuous distributions on [−1, 1] whereas the components of patterns ξp+1, . . . , ξP have discrete distributions on [−1, 1] with no atom at 0. We show that metastable behaviour occurs if there is at least one pattern of each type and 2p+7 < P . In this case we provide sharp asymptotics on metastable exit times and the corresponding capacities. Our main result can be applied to generalized Hopfield models with any values of 0 ≤ p ≤ P by adding auxiliary patterns and choosing a potential which depends only on the original patterns and coincides with the original potential. We employ the potential theoretic approach developed by Bovier et al. and use an analysis of the discrete Laplacian in the space of spin configurations to obtain lower bounds on capacities. Moreover, we include the possibility of multiple saddle points with the same value of the rate function and the case that the energy surface is degenerate around critical points.

برای دانلود متن کامل این مقاله و بیش از 32 میلیون مقاله دیگر ابتدا ثبت نام کنید

ثبت نام

اگر عضو سایت هستید لطفا وارد حساب کاربری خود شوید

منابع مشابه

محاسبه ظرفیت شبکه عصبی هاپفیلد و ارائه روش عملی افزایش حجم حافظه

The capacity of the Hopfield model has been considered as an imortant parameter in using this model. In this paper, the Hopfield neural network is modeled as a Shannon Channel and an upperbound to its capacity is found. For achieving maximum memory, we focus on the training algorithm of the network, and prove that the capacity of the network is bounded by the maximum number of the ortho...

متن کامل

High-dimensional inference with the generalized Hopfield model: principal component analysis and corrections.

We consider the problem of inferring the interactions between a set of N binary variables from the knowledge of their frequencies and pairwise correlations. The inference framework is based on the Hopfield model, a special case of the Ising model where the interaction matrix is defined through a set of patterns in the variable space, and is of rank much smaller than N. We show that maximum like...

متن کامل

Pii: S0893-6080(99)00042-8

This paper presents a theoretical analysis on the asymptotic memory capacity of the generalized Hopfield network. The perceptron learning scheme is proposed to store sample patterns as the stable states in a generalized Hopfield network. We have obtained that …n 2 1† and 2n are a lower and an upper bound of the asymptotic memory capacity of the network of n neurons, respectively, which shows th...

متن کامل

Building a Telescope to Look Into High-Dimensional Image Spaces

In Grenander’s work, an image pattern is represented by a probability distribution whose density is concentrated on different low-dimensional subspaces in the highdimensional image space. Such probability densities have an astronomical number of local modes corresponding to typical pattern appearances. Related groups of modes can join to form macroscopic image basins (known as Hopfield memories...

متن کامل

The number of metastable states in the generalized random orthogonal model

We calculate the number of metastable states in the generalized random orthogonal model. The results obtained are verified by exact numerical enumeration for small systems sizes but taking into account finite size effects. These results are compared with those for Hopfield model in order to examine the effect of strict orthonormality of neural network patterns on the number of metastable states...

متن کامل

ذخیره در منابع من


  با ذخیره ی این منبع در منابع من، دسترسی به آن را برای استفاده های بعدی آسان تر کنید

برای دانلود متن کامل این مقاله و بیش از 32 میلیون مقاله دیگر ابتدا ثبت نام کنید

ثبت نام

اگر عضو سایت هستید لطفا وارد حساب کاربری خود شوید

عنوان ژورنال:

دوره   شماره 

صفحات  -

تاریخ انتشار 2009