Perturbative M-Sequences for Auditory Systems Identification

Mark Kvale, Christoph E. SchreinerNeurIPS 1997

Overview

Researchers have developed a "perturbative" method using m-sequences to map how the brain processes sound while it is already busy listening to complex stimuli like speech. By adding a tiny, mathematically "clean" nudge to a base sound, they can extract the linear response of neurons at specific operating points. This approach reveals that the auditory system's behavior is highly contextual, shifting significantly depending on the background acoustic environment.

The Challenge of Biological System Identification

Identifying how biological systems—like the auditory cortex—transform inputs into outputs is notoriously difficult. Traditional "systems identification" often relies on Volterra or Wiener expansions, which model behavior as a series of polynomials. However, these methods face two major hurdles in biology:

  1. Numerical Instability: They frequently break down when dealing with highly nonlinear systems.
  2. Adaptation: Biological systems are "chameleons." A neuron might behave linearly when exposed to white noise but completely change its processing strategy when exposed to natural sounds or speech.

Researchers need to probe the system's "local" behavior (its linear response) without the probe itself changing the system's state. The goal is to understand how a neuron handles a small change while it is already tuned into a specific "operating point," such as a specific modulation depth or a background conversation.

Methods: The Perturbative Twist

To solve this, the authors introduce perturbative m-sequences.

An m-sequence is a binary pseudo-random sequence (+1s and -1s) that looks like noise but possesses a special mathematical property: its autocorrelation is nearly a delta function (an impulse).

Instead of using the m-sequence as the only stimulus, the researchers add a very small amount of it to a base stimulus. Because the m-sequence is mathematically "orthogonal" to the background noise over time, cross-correlating the neuron's response with the original m-sequence allows researchers to extract the "kernel" (the receptive field) specific to that background stimulus.

Waveforms illustrating the base sinusoid, the m-sequence, and the combined perturbative stimulus.

Furthermore, because neurons in the Inferior Colliculus (IC) are tuned to the envelope of sound rather than just raw pressure, the authors applied the m-sequence to the amplitude modulation (AM) of a carrier tone. This allowed them to measure the Modulation Transfer Function (MTF) at different operating points of modulation depth.

Experimental Setup

The study validated this method using experimental data from anesthetized cats, targeting single-unit extracellular recordings from the central nucleus of the Inferior Colliculus (IC). Out of 20 recorded units, 10 provided stable enough data for full analysis. The stimuli consisted of m-sequences of length $2^{15}-1$ (32,767 steps) presented over 10 trials per unit.

Key Insights and Results

The "Nudge" Advantage and System Stability

Unlike impulse stimuli—which act like hitting a system with a hammer—m-sequences spread the probing energy over time. This keeps the signal-to-noise ratio high while ensuring the "nudge" stays very close to the original operating point, leading to more accurate local measurements.

Crucially, the firing rates for a pure sinusoid and a sinusoid plus the m-sequence perturbation remain nearly identical. This proves the method is truly "perturbative"—it probes the system without disrupting the neuron's overall state.

Firing rates for a pure sinusoid versus a sinusoid with an m-sequence perturbation, demonstrating system stability.

Context Matters: Kernel Evolution

The linear response of a neuron is not a fixed property. The study found that as the "background" modulation depth changes, the neuron's response kernel changes shape. At low modulation depths, the temporal kernels showed "triphasic" behavior (three distinct peaks/valleys). As modulation depth increased, this shifted to "monophasic" behavior.

Using Fast Fourier Transforms (FFT) on the kernels, the authors also showed that the bandwidth of the neuron's modulation transfer function broadens as the modulation depth increases.

Nonlinearities are Significant

The researchers demonstrated that nonlinear responses in the presence of sinusoidal modulations are nearly as large as the linear responses. This means that a simple MTF, which assumes linearity, provides an incomplete picture of how a neuron actually works in complex environments.

Limitations and Future Work

  • Systematic Noise: The primary constraint is the trade-off between the duration of the experiment and the size of the perturbation. Extracting a clean signal from a tiny perturbation requires long recording times, which is challenging with live biological subjects.
  • Phase-Locking Requirement: This specific method is most effective for systems where units "phase-lock" (fire in sync) to a stimulus.
  • Higher-Order Kernels: While this paper focused on linear responses, the authors note that with a lower noise floor, the same m-sequence data could potentially be used to derive second-order (nonlinear) kernels, providing an even deeper map of neural processing.

Summary updated

BibTeX
@inproceedings{kvale1997perturbative,
  author = {Mark Kvale and Christoph E. Schreiner},
  title = {{Perturbative M-Sequences for Auditory Systems Identification}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {1997},
}
Figure