Visual shape discrimination in goldfish, modelled with the neural circuitry of optic tectum and torus longitudinalis.

The ability of fish to discriminate shapes visually is well documented. What is not understood are the neural mechanisms employed by fish and other vertebrates that lack a cerebral cortex to distinguish even simple geometric patterns. All the behavioral and anatomical data in fishes point to the midbrain optic tectum as the essential structure, but physiological studies of it have shown only circular contrast-detecting receptive fields and oriented edge detectors with nothing more elaborate that is reliable. Attempting to solve this conundrum, I built a model of tectum with neuron-like elements. When shown objects moving in space, the model forms an attentional locus that tracks one object on a retinotopic layer of simulated tectal pyramidal neurons. The objects elementary features, when binned together, allow it to be distinguished from other objects to some extent; it fails when objects differ only in the features spatial relationships, which fish can use for discrimination. The models solution is to bias the attentional locus to the edges of a shape in imitation of goldfish that naturally attend to the top of shapes. Redirection of the attentional locus from a shapes center is achieved by spatially offsetting synaptic inputs to the pyramidal neurons, effected by the torus longitudinalis (an elongated nucleus at the medial rim of each tectal lobe) and its prolific axonal projections to the pyramidal neurons. The models shape discrimination was compared to goldfish in the extensive behavioral data of Bowman & Sutherland (1969) who used shapes with points and projections. One test series showed that fish were sensitive to the relative number of points on the tops of shapes. In another, fish could be trained to discriminate points on the sides. By using different offset connections and only one elementary feature detector for small dark spots, the model successfully emulated the two sets of goldfish data, as judged by significant correlations between model response and fish discrimination over 21 pairs of shapes in each series.