Cortical computation of perceptual salience
Shih‐Cheng Yen · Scholarly Commons (University of Pennsylvania) · 1998
In this dissertation, we investigate the neuronal basis of the Gestalt rules of perception. We present a biologically-based neural network model of primary visual cortex capable of extracting image contours according to the Gestalt definition of perceptual salience. Our model draws on recent results in cortical anatomy, physiology and psychophysics that suggest that context dependent responses appear as early as the first cortical visual area. We also analyze the temporal response properties of model cells to explore the functional role of synchronization in neuronal assemblies. The model consists of orientation selective cortical units interconnected via long-distance horizontal connections using a connection pattern, which emphasizes smooth curvatures using constraints in both spatial position and orientation. Cortico-cortical connections allow the oriented units to facilitate each other, as has been observed physiologically (Gilbert, 1992). The facilitation triggers a bursting firing mode in the cortical units, much like the repetitive firing patterns of the chattering cells observed by Gray and McCormick (1996) in cat striate cortex. The cortical units then synchronize with other cortical units via the long-distance cortico-cortical connections. The sum of synchronized activity then provides a measure of perceptual salience. The model accounts for a number of recent psychophysical findings on contour detection, including Field, Hayes and Hess (1993), Polat and Sagi (1993, 1994), Kapadia, Ito, Gilbert and Westheimer (1995), Kovacs and Julesz (1993, 1994). The model was also tested on real grayscale images, and was shown to extract the contours that are judged most salient by human observers. Perhaps our major finding is that the properties of synchronization qualitatively account for psychophysical results showing that closure increases contour salience. This result holds regardless of the particular synchronization mechanism and seems to be a general property of the topology of open and closed chains of synchronization elements. We present both analytical and psychophysical evidence that argues that the salience of a contour is encoded in the amount of synchronized activity of the underlying neuronal population rather than the elevated firing rates of the individual cells. The results of the model provide a testable experimental paradigm for testing the functional significance of synchronization.