A Bio-Inspired Sound Localization Spiking Neural Network with Unsupervised Local Plasticity and Proximity Learning

ICLR 2026 Conference Submission17134 Authors

Keywords: unsupervised learning, spike-timing-dependent plasticity (STDP), spiking neural network (SNN), neuromorphic system, self-organization, sound source localization (SSL)

TL;DR: This paper proposes a new unsupervised local learning principle that leverages biological mechanisms for efficient learning; its integration with a bio-inspired SNN demonstrates a record-high accuracy with a human-level resolution.

Abstract: In this paper, we propose an unsupervised learning principle that leverages the neuro-inspired local plasticity and biophysiological characteristics of the brain for the learning of spiking neural networks (SNNs) without labels. The learning principle synergistically combines morphological features and biochemical phenomena in the brain cortex, guiding networks to self-organize their connectivity without global error backpropagation. The learning principle is based on two local plasticity rules. One is latency-mediated spike timing-dependent plasticity, formulated by combining the original STDP with axonal latency. The other is proximity learning, mediated by the volume transmission of neurotransmitters among neurons. We successfully applied these plasticity rules to a spiking model of the avian auditory cortex and observed the self-organization of the network, which results in the accurate localization of sound sources. After being trained using interaural time difference (ITD)-encoded spike trains, the network converged to synaptic connectivity resembling the famous Jeffress model. The performance evaluation results presented demonstrate that the proposed learning principle enables the SNN to localize sound sources with accuracy and resolution higher than those achieved by supervised learning rules.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 17134