A wide variety of mathematical and structural models that have been proposed over the past 35 years or so to describe some of the feature extraction operations found in arthropod compound eyes and vertebrate retinas have been examined. The detection or preferred response to an object's shape may in fact occur as the result of signal processing by the neural equivalent of a spatial filter. Fukushima illustrated how these operations could occur as visual signals propagate from one layer to the next of his static, visual pattern feature extractor.
Spatiotemporal filtering to select shape and velocity properties of a moving visual object was advanced by Zorkoczy, who used Boolean (logic) elements in arrays underlying the receptors, and also signal delays. It is the delay elements that enable his filters to respond to moving objects.
The specialized property of insect optomotor behavior (walking or in flight) has given rise to several models for the visual-to-motor reponse system responsible. A number of workers (Reichardt, Thorson, Bliss, Chapple, Northrop) have examined the turning tendency of walking insects, the torque on the neck, and neck muscle motor nerve activity as an insect tries to turn its head to follow the moving stripes. These optomotor outputs vary in intensity with stripe velocity, period, and contrast. Reichardt proposed an analog signal-processing model that used the outputs of two receptors (or pairs of receptors over the whole compound eye). This chapter examined Reichardt's simple DDC correlation model mathematically and derived an expression for its output response.
Requirements for and properties of speculative neural matched filters were examined. A matched filter in optics gives a maximum output (at one point) when presented with the object to be detected in noise. A neural matched filter would presumably to the same. The template filter for the object to be detected would have to be stored in the optic lobes, perhaps as some Fukushima-type filter based on synaptic connectivities between layers of neuropile.
ANNs, like any radically new technology, had a slow start in their development and applications. Three early ANNs whose designs were motivated by visual system feature extraction operations were examined. The purpose of ANNs is not to model BNNs, but to perform various pattern recognition operations with diverse applications. The salient feature of an ANN is that it can be trained, i.e., it "learns" to perform its discrimination tasks. The concept of a trainable BNN goes back to the seminal work of Hebb (1949). Hebb, a physiological psychologist, postulated that a BNN could learn some discrimination task by a purposeful readjustment of its synaptic connections. All ANNs have training algorithms that adjust the weights (transmission gains) between layers of signal nodes; such adjustments are analogous to Hebbian synaptic changes. Modern ANNs are not very "neural" in their architecture of operation; their roots were neural, however.
Was this article helpful?
This guide will help millions of people understand this condition so that they can take control of their lives and make informed decisions. The ebook covers information on a vast number of different types of neuropathy. In addition, it will be a useful resource for their families, caregivers, and health care providers.