Generalization in visual recognition by the honeybee (Apis mellifera): A review and explanation

Abstract

During a century of studies on honeybee vision, generalization was the word for the acceptance of an unfamiliar pattern in the place of the training pattern, or the ability to learn a common factor in a group of related patterns. The ideas that bees generalize one pattern for another, detect similarity and differences, or form categories, were derived from the use of the same terms in the human cognitive sciences. Recent work now reveals a mechanistic explanation for bees. Small groups of ommatidia converge upon feature detectors that respond selectively to certain parameters that are in the pattern: modulation in the receptors, edge orientations, or to areas of black or colour. Within each local region of the eye the responses of each type of feature detector are summed to form a cue. The cues are therefore not in the pattern, but are local totals in the bee. Each cue has a quality, a quantity and a position on the eye, like a neuron response. This summation of edge detector responses destroys the local pattern based on edge orientation but preserves a coarse, sparse and simplified version of the panorama. In order of preference, the cues are: local receptor modulation, positions of well-separated black areas, a small black spot, colour and positions of the centres of each cue, radial edges, the averaged edge orientation and tangential edges. A pattern is always accepted by a trained bee that detects the expected cues in the expected places and no unexpected cues. The actual patterns are irrelevant. Therefore we have an explanation of generalization that is based on experimental testing of trained bees, not by analogy with other animals. Historically, generalization appeared when the training patterns were regularly interchanged to make the bees examine them. This strategy forced the bees to ignore parameters outside the training pattern, so that learning was restricted to one local eye region. This in turn limited the memory to one cue of each type, so that recognition was ambiguous because the cues were insufficient to distinguish all patterns. On the other hand, bees trained on very large targets, or by landing on the pattern, learned cues in several eye regions, and were able to recognize the coarse configural layout.