Revealing Causal Graph Structure in Pigeon Flocks Using Attentions in Temporal ConvNet
Document Type
Conference Proceeding
Publication Date
1-1-2023
Abstract
The study of collective animal behavior has long been of interest to researchers across various disciplines. In this study, a modified version of the Attention-based Dilated Depthwise Separable Temporal Convolutional Networks (AD-DSTCN) architecture was used to analyze high-resolution spatiotemporal data of moving pigeons with the goal of producing causal graph structures in pigeon flocks. The modification allowed for causality inference on data where individuals have multidimensional feature representations. The study revealed several important findings. Firstly, the results showed that, although weak external influences may still be present, each pigeon primarily relies on its own knowledge and decision-making. Secondly, the results also indicated that edges in the graph correspond to pairs of pigeons that are close to each other, suggesting that pigeons that are physically near each other tend to have stronger social connections. Thirdly, the study suggested that there is no single persistent pigeon that holds the most influence in the flock. Instead, the most influential node appears to shift at various time intervals, a concept known as intermittent switching. These findings offer new insights into the complexity of animal behavior and its underlying mechanisms, as well as a deeper understanding of how pigeons make decisions and interact within a flock. Ultimately, the use of the modified AD-DSTCN architecture offers a promising approach for understanding collective animal behavior.
Recommended Citation
J. P. Layacan, A. M. Donaire, H. Felipe and C. K. Go, "Revealing Causal Graph Structure in Pigeon Flocks Using Attentions in Temporal ConvNet," 2023 International Joint Conference on Neural Networks (IJCNN), Gold Coast, Australia, 2023, pp. 1-7, doi: 10.1109/IJCNN54540.2023.10191157.
