Data_Sheet_4_Numerosity Categorization by Parity in an Insect and Simple Neural Network.CSV

A frequent question as technology improves and becomes increasingly complex, is how we enable technological solutions and models inspired by biological systems. Creating technology based on humans is challenging and costly as human brains and cognition are complex. The honeybee has emerged as a valuable comparative model which exhibits some cognitive-like behaviors. The relative simplicity of the bee brain compared to large mammalian brains enables learning tasks, such as categorization, that can be mimicked by simple neural networks. Categorization of abstract concepts can be essential to how we understand complex information. Odd and even numerical processing is known as a parity task in human mathematical representations, but there appears to be a complete absence of research exploring parity processing in non-human animals. We show that free-flying honeybees can visually acquire the capacity to differentiate between odd and even quantities of 1–10 geometric elements and extrapolate this categorization to the novel numerosities of 11 and 12, revealing that such categorization is accessible to a comparatively simple system. We use this information to construct a neural network consisting of five neurons that can reliably categorize odd and even numerosities up to 40 elements. While the simple neural network is not directly based on the biology of the honeybee brain, it was created to determine if simple systems can replicate the parity categorization results we observed in honeybees. This study thus demonstrates that a task, previously only shown in humans, is accessible to a brain with a comparatively small numbers of neurons. We discuss the possible mechanisms or learning processes allowing bees to perform this categorization task, which range from numeric explanations, such as counting, to pairing elements and memorization of stimuli or patterns. The findings should encourage further testing of parity processing in a wider variety of animals to inform on its potential biological roots, evolutionary drivers, and potential technology innovations for concept processing.

随着技术持续进步且日趋复杂，一个广受关注的问题是：我们该如何打造受生物系统启发的技术解决方案与模型。以人类为参照开发技术极具挑战性且成本高昂，这是因为人类大脑与认知机制极为复杂。蜜蜂已成为一种极具价值的比较模型，其展现出若干类认知行为。相较于大型哺乳动物的大脑，蜜蜂大脑的结构相对简易，这使得它们能够完成分类等学习任务，而这类任务可通过简单神经网络进行模拟。对抽象概念进行分类，是我们理解复杂信息的关键环节。在人类的数学表征体系中，奇偶数值处理被称为奇偶判断任务（parity task），但目前尚未有针对非人类动物的奇偶判断处理的相关研究。本研究证实，自由飞行的蜜蜂可通过视觉学习，掌握区分1至10个几何元素的奇偶数量的能力，并能将该分类能力外推至11和12这两个全新的数量区间，这表明相对简易的神经系统也可实现此类分类任务。基于上述发现，我们构建了一个包含5个神经元的神经网络，该网络可可靠地对最多40个几何元素的数量进行奇偶分类。尽管该简单神经网络并未直接基于蜜蜂大脑的生物学机制构建，但其构建目的是验证简易系统能否复现我们在蜜蜂身上观测到的奇偶分类结果。因此，本研究表明，此前仅在人类身上观测到的任务，可被拥有相对较少神经元的大脑完成。我们还探讨了蜜蜂能够完成此类分类任务的可能机制与学习过程，其涵盖了从计数等数值解释，到配对元素与记忆刺激或模式等多种可能性。本研究结果将推动在更多种类的动物中开展奇偶判断处理的相关测试，以阐明该能力潜在的生物学根源、进化驱动因素，以及用于概念处理的潜在技术创新方向。