Understanding the significance of specific communication cues necessitates our ability to detect them, particularly with transient frequency modulation signals like chirps produced by an electric organ discharge in electric fish. Previous research has mainly focused on immobilizing or artificially stimulating fish or physically separating them, conditions unfavorable for natural communication.
To address this challenge, I designed a convolutional neural network-based detector capable of detecting chirps in freely behaving fish. Despite initially training the model on simulated data, it surprisingly performed well on real-world recordings after some fine-tuning. Using this version, I successfully detected approximately 50,000 chirps, marking the largest dataset at that time.
The following image illustrates a short segment of a recording featuring two fish. Chirps are visible as frequency fluctuations from the baseline of one of the two fish on a spectrogram. The dots indicate where the detector identified a chirp.
Performing preliminary analyses utilizing this detector, we discovered that chirps could potentially be utilized by the losing fish to indicate submission during competition for a shelter among two fish.
A first try of detecting the transient communication signals of weakly electric fish on a spectrogram using a convolutional neural network.