Red de Hopfield
Una Red Hopfield es una forma de red recurrente red neuronal artificial that serves primarily as an associative memory system. Introduced by John Hopfield in 1982, this network is designed to store and recall basada en un modelo de memoria direccionable por contenido.
El architecture of a Hopfield Network consists of a single layer of interconnected neurons, where each neuron is connected to every other neuron, but not to itself. These connections have weights that can be either positive or negative, allowing the network to represent various patterns. The key feature of a Hopfield Network is its ability to converge to a stable state that represents a stored pattern when it is presented with a partial or noisy version of the input.
Durante su funcionamiento, la red actualiza el estado de sus neuronas de manera asincrónica o sincrónica en función de la suma ponderada de sus entradas, aplicando una función umbral para determinar si cada neurona debe activarse (es decir, emitir un 1) o permanecer inactiva (emitir un 0). Los patrones almacenados se representan como mínimos locales en el paisaje de energía de la red, y el proceso de recordar un patrón puede verse como la red minimizando su energía al moverse hacia estos mínimos.
Hopfield Networks are particularly valuable for tasks such as error correction, pattern recognition, and optimization problems. However, they have limitations, including a capacity constraint on the number of patterns they can reliably store, which is approximately 0.15 times the number of neurons in the network. Despite these limitations, Hopfield Networks laid the groundwork for further research in redes neuronales y que inspiró el desarrollo de modelos más complejos.