In this proposal, we examine the applicability of the selective amplification principle to signal processing in the auditory cortex. This principle is based on a model developed recently in VI [.Koch Douglas 1995.] that depends on selective amplification of the feedforward sensory input. The selective amplification arises from the cooperative interaction of a small population of cortical neurons, in which recurrent excitatory connections provide amplification of the neurons' response to the feedforward input. The population restores cooperatively the incoming stimulus towards a pattern that is latent in the recurrent connectivity, and thereby permits meaningful outputs to be extracted from incomplete or noisy input patterns. This restoration is related to recall in content addressable (associative) memories composed of simple linear-threshold neurons. However, those networks typically have very strong positive feedback and many stable attractors, while the cortical circuits presented here operate in a domain where they can represent in a proportional manner various aspects of the input, such as its contrast or velocity. Thus, the cortical circuits combine aspects of analog (``smart'' amplification) with digital (signal restoration) signal processing.
We plan to explore whether similar circuits explain receptive field properties in the auditory areas. Even in the primary sensory cortical areas, the high degree of cortical interconnectivity, compared to the small number of extracortical inputs, raises the possibility that receptive field properties are not so much determined by the specific patterns of thalamic afferents, but are shaped by the collective behavior of large populations of cortical cells, making cortex a substantially richer modifiable architecture than standard feedforward models. We have extended the selective amplification principle to account for operations such as co-ordinate transformation, solution of stereocorrespondence, and a form of hypothesis satisfaction that could be applied to a broad class of pattern recognition problems such as those discussed in Thrust area V [. Mahowald motion 1996, Hahnloser 1996.].