Abstract

The well-being of mice is critical in pre-clinical research laboratory studies to achieve high-quality experimental results and to mitigate ethical concerns. In this context, the paper proposes a generative intelligent solution for producing human-readable descriptions of mice’s behavior living in a video sensorized cage. By deploying models on off-the-shelf edge devices, this work contributes to the current efforts of the EdgeAI Foundation aimed at bringing Generative AI at the edge. The neural models employed were built using the VideoMAE transformer as an encoder, the Open Pre-Trained Transformer, as the decoder, generating textual descriptions of the predicted behavior. The tiniest version devised, named TinyV2A, achieved a 271 ms inference time and required 1.3 GB of memory on the Raspberry Pi 5. Tiny Text to Speech completed the pipeline, known as Piper, with 15.65 M parameters.

Fiorenza, G., Pau, D.P., Schettini, R. (2025). Action Prediction with Edge Generative AI for Mice Pre-clinical Studies. In: Arabnia, H.R., Deligiannidis, L., Shenavarmasouleh, F., Amirian, S., Ghareh Mohammadi, F. (eds) Computational Science and Computational Intelligence. CSCI 2024. Communications in Computer and Information Science, vol 2502. Springer, Cham. https://doi.org/10.1007/978-3-031-94937-1_1