Recognition and correct classification of visually acquired objects is important for the survival of all creatures. Classification errors almost always lead to loss of fitness: incorrect classification of another animal as a competitor may lead to costly agonistic behaviour, and the incorrect classification of a predator may lead to death. We studied classification of faces in a species of coral reef fish, P. amboinensis, which has individually distinct facial patterns, using a combination of behaviour and machine vision techniques.

A series of experiments relying on associative learning was conducted to test the facial discrimination ability of the fish both within and between species. Fish received a food reward when they successfully selected their trained stimulus (face discrimination), or a certain stimulus class (conspecific / heterospecific face). Stimuli consisted of greyscale images of the head with facial markings or of the extracted patterns alone and, depending on the experiment, were either laminated printouts or presented on a computer screen.

We found that the fish quickly learn to discriminate between two facial patterns stemming from either conspecific or heterospecific fish and that they are highly sensitive to slight pattern differences such as are found on individual conspecifics, or artificially induced differences by morphing between pairs of faces. We also found that following initial training with four pairs of conspecific / heterospecific faces, the fish were able to classify novel conspecific and heterospecific faces with 75% accuracy. Evidence for categorical perception was found when morphed facial patterns were used.

Our results show that face discrimination in fish has many similarities with what we know about human face discrimination, despite their lack of a neocortex. Machine vision techniques were employed to examine the very large space of candidate decision features and the outcomes of this analysis will be discussed.