Paper accepted for publication in Advanced Intelligent Systems (12th May 2026)
Our work on interpreting echolocation-inspired neural networks has been accepted for publication in Advanced Intelligent Systems.
This work develops a way to make AI/ML models for acoustic sensing more understandable by revealing which echo patterns they use to make decisions and how those patterns shape their predictions. By using an ensemble of shallow CNNs, the framework makes it possible to trace model decisions back to the original acoustic input, while also demonstrating that the networks rely on meaningful and robust features that transfer successfully from simulated echoes to real-world measurements. This work could lay the foundation to develop robust, transparent robotic perception systems, as well as to understand the echolocation abilities of mammals such as dolphins and bats.
This paper builds on our prior work published in the Journal of Sound and Vibration and IEEE Access, where we introduced class-specific CNN ensembles and demonstrated the ability to recognize perceptually similar shapes from acoustic echoes, in both synthetic and real environments.