Who’s that rat? Setting the score for unmarked rat identification with Deep Learning

The ability to identify unmarked animals across time and sessions is a major challenge in laboratory experiments. Researchers currently approach animal identification with Deep Learning, but previous studies suffer from methodological limitations, including insufficient statistical validation, inadequate exploration of image and learning parameters, and a lack of reproducibility-highlighting the need for a more rigorous assessment. We explore the effectiveness of Deep Learning techniques for identifying laboratory rats by training and testing Convolutional Neural Networks and other models on a dataset of 1.44 million images of 16 rats under controlled conditions. This study establishes the baseline for unmarked animal identification in behavioural research. We evaluate the scalability of these techniques and examine how image properties and learning parameters influence their performance. Our results show that the identification accuracy can exceed 90\% for a moderate number of individuals, without using additional knowledge such as heuristics or body part information. However, they also challenge widely accepted claims in the scientific literature that Deep Learning techniques can reliably distinguish up to 100 unmarked animals in experimental settings.