Interpretable Anomaly Detection in a Hippocampal Formation-Inspired Spatial Cognition Model

Takeshi Nakashima, Akira Taniguchi, Tadahiro Taniguchi, Hiroshi Yamakawa

Abstract: Interpretability is a key requirement in deploying AI systems for safety-critical tasks such as anomaly detection. Brain-inspired AI offers a promising path toward interpretable models by aligning internal representations with known structures and functions of the brain. To address the challenge of developing interpretable anomaly detection systems, we propose a novel approach based on a spatial cognition model inspired by the hippocampal formation. Our model introduces three anomaly scores, each corresponding to a probabilistic variable associated with a specific subregion of the hippocampal formation: the lateral entorhinal cortex, the medial entorhinal cortex, and the cornu ammonis-1, respectively. Experiments in a simulated home environment, involving different types of anomalies such as missing observation data, pose estimation failures, and contextual anomalies (e.g., the kidnapped robot problem), show that some of the anomaly scores are most effective in detecting the type of anomaly that activates the corresponding hippocampal formation subregion in actual neuroscience studies. These results suggest that our hippocampal formation-inspired model is capable of anomaly detection and offers insights into the functional localization of internal processing that mirrors neuroscientific findings, thereby enhancing interpretability. Our approach lays the groundwork for integrating neuroscientific principles into designing more transparent and trustworthy AI systems.

External IDs:doi:10.1007/978-981-95-4384-7_35