A model of frequency discrimination with optimal processing of auditory nerve spike intervals.

This paper investigates phase-lock coding of frequency in the auditory system. One objective with the current model was to construct an optimal central estimation mechanism able to extract frequency directly from spike trains. The model bases estimates of the stimulus frequency on inter-spike intervals of spike trains phase-locked to a pure tone stimulus. Phase-locking is the tendency of spikes to cluster around multiples of the stimulus period. It is assumed that these clusters have Gaussian distributions with variance that depends on the amount of phase-locking. Inter-spike intervals are then noisy measurements of the actual period of the stimulus waveform. The problem of estimating frequency from inter-spike intervals can be solved optimally with a Kalman filter. It is shown that the number of inter-spike intervals observed in the stimulus interval determines frequency discrimination at low frequencies, while the variance of spike clusters dominates at higher frequencies. Timing information in spike intervals is sufficient to account for human frequency discrimination performance up to 5000 Hz. When spikes are available on each stimulus cycle, the model can accurately predict frequency discrimination thresholds as a function of frequency, intensity and duration.

Shape encoding: a biologically inspired method of transforming boundary images into ensembles of shape-related features.

A biologically inspired method to encode the shape of enclosed image regions is presented. Shape encoding units, called "shape cells", are algebraic filters which receive two-dimensional (2-D) reconstructed image boundaries as input features. Shape cell outputs are calculated as three components. First, a compact description of an image boundary shape surrounding a particular shape cell is obtained in the form of a shape cell centered radial scanning vector. Shape cell activations are then calculated from radial scans, and finally, different shape cells that encode parts of the same shape must be grouped together in ensembles. This process is called feature binding. A process of iterative lateral inhibition is employed to condense the set of active shape cells before feature binding takes place. The output radial scanning vectors of shape cells provide compact descriptions of shape which are useful in object identification. The spatial pattern (2-D coordinates) of active shape cells can be used in object localization. With feature binding separate ensembles are created, even if neighboring shapes are only divided by weak boundaries. Besides its application in pattern recognition, shape encoding provides a possible mechanism of figure-ground separation. Artificial shape encoding is further concluded to be a suitable addendum to the existing collective model of biological vision.