BioNet-A: Ultrasonic echo representation network for target discrimination using active SONAR

Area of Science:

• Bioacoustics
• Machine Learning
• Sensory Neuroscience

Background:

• Ultrasonic echoes are crucial for object perception but pose challenges for conventional spectrogram-based models due to their short duration and frequency structure.
• Existing biomimetic systems often fail to optimally process these brief echoes or impose unsuitable frequency shift-invariance.

Purpose of the Study:

• To introduce BioNet5-A, a novel biomimetic encoder specifically designed to optimize the processing of ultrasonic echoes.
• To improve object discrimination and robustness in biosonar-inspired sensing applications.

Main Methods:

• Developed BioNet5-A from an autoencoder pretrained on bat vocalizations.
• Incorporated spectrotemporal attention, multi-sized convolution modules, and a symmetric encoder-decoder architecture.
• Evaluated performance on a controlled ultrasonic dataset with varying objects and noise conditions.

Main Results:

• BioNet5-A demonstrated superior performance compared to auditory spectrograms, cortical wavelets, and the conventional BioNet.
• Achieved improved echo clustering, discrimination, and robustness against noise.
• Generated compact and interpretable representations consistent with bat midbrain physiology.

Conclusions:

• BioNet5-A represents a significant advancement in processing ultrasonic echoes for object perception.
• The model offers a practical and effective front-end solution for biosonar-inspired sensing and other ultrasonic applications.