A Turing Test for Collective Motion

A widespread problem in biological research is assessing whether a model adequately describes some real-world data. But even if a model captures the large-scale statistical properties of the data, should we be satisfied with it? We developed a method, inspired by Alan Turing, to assess the effectiveness of model fitting. We first built a self-propelled particle model whose properties (order and cohesion) statistically matched those of real fish schools. We then asked members of the public to play an online game (a modified Turing test) in which they attempted to distinguish between the movements of real fish schools or those generated by the model. Even though the statistical properties of the real data and the model were consistent with each other, the public could still distinguish between the two, highlighting the need for model refinement. Our results demonstrate that we can use ‘citizen science’ to cross-validate and improve model fitting not only in the field of collective behavio...生物学研究领域普遍存在的一项挑战，便是评估某一模型能否充分刻画真实世界的某类数据。但即便模型能够捕捉数据的宏观统计特性，我们是否就能就此满足？为此，我们受艾伦·图灵（Alan Turing）的启发，开发了一种用于评估模型拟合效果的方法。我们首先构建了一个自驱动粒子模型（self-propelled particle model），其有序性与聚集性的统计特征与真实鱼群一致。随后我们邀请公众参与一款在线游戏——这是一种改良版的图灵测试（Turing test）——参与者需要辨别真实鱼群的运动轨迹，与模型生成的鱼群运动轨迹之间的差异。尽管真实数据与模型的统计特性彼此相符，但公众仍能够分辨二者的区别，这凸显了对模型进行优化完善的必要性。我们的研究结果表明，我们可以借助“公民科学（citizen science）”开展交叉验证并提升模型拟合效果，这不仅适用于集体行为...