Associative Memories in Infinite Dimensional Spaces
Neural Processing Letters
| Hi-index | 0.00 |
A neural network model of associative memory is presented which unifies the two historically more relevant enhancements to the basic Little-Hopfield discrete model: the graded response units approach and the stochastic, Glauber-inspired model with a random field representing thermal fluctuations. This is done by casting the retrieval process of the model with graded response neurons, into the framework of a diffusive process governed by the Fokker-Plank equation, which leads to a Langevin system describing the process at a microscopic level, while the time evolution of the probability density function is governed by a multivariate Fokker Planck equation operating over the space of all possible activation patterns. The present unified approach has two notable features: (i) greater biological plausibility and (ii) ability to escape local minima of energy (associated with spurious memories), which makes it a potential tool for those complex optimization problems for which the previous models failed.