Bio-inspired Working Memory for Online Auditory Pattern Drift Detection

Zhongju Yuan, Geraint A. Wiggins, Dick B.M. Botteldooren

Keywords: bio-inspired network, auditory working memory, oscillatory dynamic system, selective attention, online drift detection, unsupervised learning

Abstract: Recent advances in Audio Language Models (ALMs) have attracted unprecedented attention. However, transformer-based ALMs face challenges in long-form audio understanding due to inefficient attention allocation. To address this, we introduce a biologically inspired working memory module, BioWM (Bio-inspired Working Memory), which leverages unsupervised online drift detection as an adaptive attention allocation strategy. BioWM detects auditory pattern drifts by monitoring energy fluctuations induced by spatio-temporal shifts, enabling the model to focus on salient changes. The BioWM does not require long-term historical data or offline pretraining; instead, it adapts online with only a few steps of threshold adjustment. Our approach captures novel events while remaining robust to transient perturbations. Furthermore, BioWM exhibits oscillatory frequency-band dynamics that resemble cortical activity during working memory tasks, thereby strengthening its biological plausibility. We present comprehensive experiments demonstrating the effectiveness of BioWM and provide visualizations of its evolving internal states to highlight both performance gains and interpretability.

Primary Area: applications to neuroscience & cognitive science

Submission Number: 1566