P329 The effect of overfitting on spatial perception and flight trajectories in pigeons
Margarita Zaleshina*1, Alexander Zaleshin2
1Moscow Institute of Physics and Technology, Moscow, Russia 2Institute of Higher Nervous Activity and Neurophysiology, Moscow, Russia
*Email: zaleshina@gmail.com
Introduction
Problems of overfitting in trained systems concern not only artificial neural networks, but also living organisms and humans. Pre-trained templates can reduce processing time, but increase errors in real dynamic situation. Conventional models often use not new, but current templates with distortion, addition, prolongation. Due to overfitting data can be misinterpreted, relevant data can be filtered out [1].
In our work we study overfitting in pigeon flights. These birds often use accumulated knowledge and route-finding algorithms (guided by beacons, long roads, loft-like buildings) [2]. EEG activities in a familiar situation differ from brain activity in new conditions, it can be observed with Neurologgers and GPS trackers [3]. Methods We compared GPS tracks and brain activity of untrained and trained pigeons flying over landscapes with different information loads: near sea coast, over rural or urbanized areas. Source materials were selected from the Dryad Digital Repository and Movebank Data Repository. We calculated brain frequencies and their changes; standard deviation from average flight path; frequency of surveying (loops in trajectories); percentage of detectable "points of interest" (Fig. 1). Spatial analysis of GPS tracks, detection of landscape boundaries and detection of special points were performed using the QGIS. To identify overfitting, we computed a decrease in the flexibility of individual flights and a decrease in the power of high-frequency EEG. Results Brain activity was most pronounced near the loft and least pronounced when pigeons flew along known routes along homogeneous terrains or extended objects. Additionally, high brain activity and surveying were demonstrated by pigeons when examining points of interest or when moving from one type of landscape to another, even by trained pigeons. Trained pigeons more often preferred to fly along known track, even if it differed from the shortest route. In overfitting flights, surveying and standard deviations from the average flight track and changes in flight direction were minimal. Overfitting flights were often observed over rural terrain, less often in the coastal zone. In flocks the frequency of overfitting cases increased. Discussion Importance of overfitting is especially significant in modern conditions of accelerated emergence and use of "big" digital data. Excessive templates and strict filters can often lead to errors or to significantly limit the variability. Usage of multilayer data sources allows to accommodate and vary different planes of view, or context basic points, which helps reduce overfitting. Studying of flight pigeons paths demonstrate relationship between external environment, chosen behavior and internal settings of trained birds. Surveying increases an ability to navigate in dynamical cases or to find interesting locations.
In future we plan to continue studying of surveying and multilayer data exchange to reduce the overfitting problem.
Figure 1. Typical cases of pigeon flight and pigeon EEG-power: trained pigeon, trained pigeon, pigeon near the point of interest, pigeon after overfitting Acknowledgements - References 1. Zaleshina, M. & Zaleshin, A. (2024). Spatial Learning and Overfitting in Visual Recognition and Route Planning Tasks. IJCCI & NCTA. 1: 576-583. 2. Blaser, N. et al. (2013). Testing Cognitive Navigation in Unknown Territories: Homing Pigeons Choose Different Targets. Journal of Experimental Biology. 216(16):3123–31. 3. Ide, K. & Takahashi, S. (2022). A Review of Neurologgers for Extracellular Recording of Neuronal Activity in the Brain of Freely Behaving Wild Animals. Micromachines.13(9):1529.
