Search Papers | Poster Sessions | All Posters
Poster B166 in Poster Session B - Thursday, August 8, 2024, 1:30 – 3:30 pm, Johnson Ice Rink
One-shot auditory blind source separation using a novel neural network inspired by the auditory system
Patrick Abbs1, Sanchit Gupta1, Erika Schmitt1 (); 1Cambrya, LLC
The human brain can naturally identify and track individual sounds even amidst a cacophony of overlapping noises—a phenomenon known as the "cocktail party effect." However, computational algorithms and machine learning approaches struggle to perform single-channel blind source separation (BSS) of auditory signals. We present Density Networks (DNs), a novel class of recurrent neural network inspired by the auditory system that demonstrates one-shot BSS of auditory signals. DNs have artificial inner hair cells (IHCs) that connect to layers of artificial neurons with tonotopy, and feedback and feedforward inhibitory and excitatory mechanisms that facilitate plasticity and learning at multiple timescales. Each structure in the network has distinct learning rules and spontaneously coordinates with other actors to produce an emergent output. Therefore, network behavior is completely interpretable in real-time by monitored behaviors ranging from synaptic weight changes and firing rates to population-level neuronal synchrony. This biologically inspired algorithm learns and then follows new sounds within 300 milliseconds, akin to human auditory performance. DNs also outperformed two state-of-the-art single-channel BSS separation methods—improving sound separation quality by at least 160%. Unlike popular deep learning algorithms DNs are unsupervised, making them suitable for lifelong learning in real-world sensory environments.
Keywords: signal processing working memory unsupervised auditory system