A two-layer neural system for shift-invariant pattern recognition is proposed. Model neurons are endowed with physiological dynamics involving membrane potentials and axonic spikes. Synapses between the two layers are plastic and change according to spike coincidences (Hebbian rules). The first neural network (encoder network) extracts features from a presented pattern and codes the neighborhood relationship of features by coincident activity of neurons. The second network (memory network) has stored several patterns. During recognition of a presented pattern the neural system establishes a strong projection between the first and the second layer, enhances activity in the set of those neurons represent the presented patterns, and supresses activity of other neurons. Synaptic plasticity according to Hebbian rules make it possible to generate a projection which preserves feature neighborhood relationships.
|Original language||English (US)|
|Title of host publication||IEEE Int Conf on Neural Networks|
|Publisher||Publ by IEEE|
|Number of pages||8|
|State||Published - 1988|
ASJC Scopus subject areas