Insect Bio-inspired Neural Network Provides New Evidence on How Simple Feature Detectors Can Enable Complex Visual Generalization and Stimulus Location Invariance in the Miniature Brain of Honeybees

The ability to generalize over naturally occurring variation in cues indicating food or predation risk is highly useful for efficient decision-making in many animals. Honeybees have remarkable visual cognitive abilities, allowing them to classify visual patterns by common features despite having a relatively miniature brain. Here we ask the question whether generalization requires complex visual recognition or whether it can also be achieved with relatively simple neuronal mechanisms. We produced several simple models inspired by the known anatomical structures and neuronal responses within the bee brain and subsequently compared their ability to generalize achromatic patterns to the observed behavioural performance of honeybees on these cues. Neural networks with just eight large-field orientation-sensitive input neurons from the optic ganglia and a single layer of simple neuronal connectivity within the mushroom bodies (learning centres) show performances remarkably similar to a large proportion of the empirical results without requiring any form of learning, or fine-tuning of neuronal parameters to replicate these results. Indeed, a model simply combining sensory input from both eyes onto single mushroom body neurons returned correct discriminations even with partial occlusion of the patterns and an impressive invariance to the location of the test patterns on the eyes. This model also replicated surprising failures of bees to discriminate certain seemingly highly different patterns, providing novel and useful insights into the inner workings facilitating and limiting the utilisation of visual cues in honeybees. Our results reveal that reliable generalization of visual information can be achieved through simple neuronal circuitry that is biologically plausible and can easily be accommodated in a tiny insect brain.Author Summary: We present two very simple neural network models based directly on the neural circuitry of honeybees. These models, using just four large-field visual input neurons from each eye that sparsely connect to a single layer of interneurons within the bee brain learning centres, are able to discriminate complex achromatic patterns without the need for an internal image representation. One model combining the visual input from both eyes showed an impressive invariance to the location of the test patterns on the retina and even succeeded with the partial occlusion of these cues, which would obviously be advantageous for free-flying bees. We show that during generalization experiments, where the models have to distinguish between two novel stimuli, one more similar to a training set of patterns, that both simple models have performances very similar to the empirical honeybee results. Our models only failed to generalize to the correct test pattern when the distractor pattern contained only a few small differences; we discuss how the protocols employed during training enable honeybees to still distinguish these stimuli. This research provides new insights into the surprisingly limited neurobiological complexity that is required for specific cognitive abilities, and how these mechanisms may be employed within the tiny brain of the bee.