FrogAI: low-power bioacoustics on the Edge
Can we detect invasive species the moment they call, using a device no bigger than a matchbox? FrogAI is a bioacoustic monitoring system that brings deep learning out of the lab and into the field, running species classifiers directly on low-power microcontrollers and edge AI hardware. We focus on invasive amphibians that threaten the endemic fauna of Okinawa and the Ryukyu Islands, including the cane toad (Rhinella marina), the common tree frog (Polypedates leucomystax), and the spot-legged tree frog (Polypedates megacephalus). Rather than adapting off-the-shelf models, we design new neural network architectures through optimization, tailored from the ground up to the tight memory and compute budgets of embedded devices. We pair this with quantization-aware training so the models keep their accuracy after being compressed to 8-bit integers, which allows reliable inference on platforms from the Sony Spresense and Arduino Portenta to NPU-equipped Alif Ensemble devices. The project has drawn interest from local authorities as a tool for early detection and biosecurity, and we are now developing FrogAI into a deployable monitoring platform. We welcome collaborators, data partners, and organisations interested in bringing it to the field.
Zebrafish behavioral analysis in 3D
How do social hierarchies emerge from physical interactions? We address this through high-resolution 3D tracking of zebrafish aggressive behavior by tracking the head, pectoral fin, and tail of each fish throughout 5-hours recordings. We conduct 3D posture tracking by integrating two complementary, state-of-the-art approaches, SLEAP and idTracker.ai. This high-resolution 3D tracking allows novel scientific hypotheses for understanding animal behavior.
For example, wild-type zebrafish pairs exhibit a clear temporal progression in social behavior. During initial interactions (first 50 minutes), fish engage in mutual aggressive displays with symmetric gaze patterns; both fish orient toward each other with similar frequency. As dominance relationships stabilize, gaze asymmetry emerges: the subordinate fish consistently orients toward the dominant individual, while the dominant fish shows reduced attention to its subordinate. In contrast, mecp2 mutant pairs fail to establish this characteristic asymmetry, maintaining symmetric gaze patterns throughout the recording period. This suggests that normal dominance establishment requires intact mecp2 function, and that gaze asymmetry serves as a quantitative behavioral signature of hierarchy formation in zebrafish social interactions.