Artificial echolocation seeks to emulate biosonar for inferring shape (scatterer geometry) from acoustic echoes. Many promising approaches use convolutional neural networks (CNNs), but how these networks make decisions remains unclear, leaving model development largely empirical. This work introduces an interpretability framework for a robotic perception system composed of collections of specialized CNNs (SCNNs), each trained to recognize a specific shape. By leveraging the shallow SCNN architecture, convolutional features are mapped to decision neurons, and a feature‐importance metric is defined to reveal the echo cues that drive classifications. The analysis shows that SCNNs rely on the relative emphasis of several cues and that perturbing these cues predictably alters model outputs. In addition, decreasing signal‐to‐noise ratio in measured echoes is shown to obscure these cues, explaining the limits of generalization from synthetic to measured data. The framework transforms an opaque classifier into a transparent one, offering new insight into how neural networks parse echolocation echoes and providing guidance for the design of more interpretable and robust robotic perception systems.
Patil et al. (Sun,) studied this question.