A Neural Oscillator Model of Auditory Selective Attention

Stuart N. Wrigley, Guy J. BrownNeurIPS 2001

Introduction

In a crowded room, your brain performs "Auditory Scene Analysis" (ASA)—the process of separating a jumble of sound waves into distinct "streams" like a friend’s voice or background music. Traditionally, researchers thought the brain grouped sounds automatically, and attention only stepped in later to pick which group to listen to.

However, newer psychoacoustic evidence suggests that without attention, some types of auditory grouping simply don't happen. This paper introduces a computational model that treats auditory attention not just as a "spotlight" for selecting sounds, but as a fundamental "glue" required to form auditory streams in the first place. By using a network of neural oscillators and an "attentional leaky integrator," the model successfully replicates human-like performance in segregating mistuned harmonics and tracking interleaved tone sequences.

Methods and Architecture

The authors built a multi-stage pipeline that mimics the human auditory pathway to integrate both exogenous (low-level, automatic) and endogenous (high-level, conscious) attention.

Architecture of the auditory attention model.
  • Peripheral Processing: A bank of 128 gammatone filters simulates the cochlea, followed by half-wave rectification to mimic auditory nerve firing.
  • Pitch Analysis: The model generates a "correlogram" (using autocorrelation) to identify periodicities and harmonic relationships.
  • Neural Oscillator Network: The core of the model is a 1D array of 128 oscillators based on the LEGION (Locally Excitatory Globally Inhibitory Oscillator Network) framework. Oscillators that represent the same sound source synchronize their firing, while those representing different sources desynchronize. This unidimensional approach is argued to be more biologically and computationally plausible than previous 2D time-frequency grids.
  • The ALI (Attentional Leaky Integrator): This is the paper's primary innovation. It’s a specialized unit that receives input from the oscillators, modulated by a "top-down" interest vector. It acts as a temporal anchor; if an oscillator’s activity coincides with a peak in the ALI, that sound is considered to be in the "attentional foreground."
  • Age Trackers: To implement the "old-plus-new" heuristic (the tendency to hear a continuing sound as a single stream), the model uses leaky integrators that track how long a specific frequency has been active.

Key Insights and Results

As a theoretical modeling paper, it evaluates the model against classic psychoacoustic stimuli used in landmark human studies, specifically focusing on mistuned harmonic complexes and tone sequences.

Model performance on mistuned harmonics.
  • Mistuning Thresholds: The model accurately replicated human data showing that a harmonic must be mistuned by more than ~8% to be perceived as a separate sound. Below this, it contributes to the overall pitch of the complex; above it, it is "segregated" (as seen in the pitch shift comparisons with human data).
  • Attention as a Requirement for Grouping: The model demonstrates that stream formation is an emergent property of the interaction between stimulus-driven activity and attentional feedback. It reinforces the "oscillatory correlation theory"—the idea that the brain "labels" related acoustic features by making the neurons representing them fire at the exact same time.
  • The "Old-Plus-New" Heuristic: By using "age trackers," the model elegantly explains why a harmonic that starts before a complex tone is perceived as a separate object, whereas one that starts with the complex is fused into it.
  • Captor Tones and Streaming: The model successfully simulated "auditory capture," where preceding tones at a specific frequency "pull" a harmonic out of a subsequent complex tone. In tests with alternating tones, the model correctly predicted that streaming (perceiving two separate sequences) occurs when frequency separation exceeds 3 to 4 semitones, matching experimental findings.

Limitations and Future Work

  • Resolved Harmonics: The authors note that their correlogram-based approach doesn't perfectly capture the "dominance region" of pitch—the fact that human pitch perception is more heavily influenced by lower-frequency, "resolved" harmonics.
  • Spatial Cues: The current model is monaural (one ear). The authors suggest that future versions should include binaural effects, such as "inter-ear attentional competition," to better simulate how we use spatial location to filter sounds.
  • Complexity: While successful on synthetic tones, it remains to be seen how this oscillator framework scales to highly non-linear, non-periodic sounds like speech in a real-world "cocktail party" environment.

Summary updated

BibTeX
@inproceedings{wrigley2001neural,
  author = {Stuart N. Wrigley and Guy J. Brown},
  title = {{A Neural Oscillator Model of Auditory Selective Attention}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2001},
}
Figure