Improving the entropy estimate of neuronal firings of modeled cochlear nucleus neurons

Andrea Grigorescu, Marek Rudnicki, Michael Isik, Werner Hemmert, Stefano RiniInterspeech 2012

TL;DR

This paper introduces a novel time-varying and time-dependent entropy estimation method to better understand how cochlear nucleus neurons encode speech. By applying this to simulated neuronal responses, they reveal that these neurons exhibit high entropy rates, reflecting their precise temporal coding of sound features, and demonstrate how this method can reveal neuronal memory.

What Problem Does This Paper Address?

Understanding how the brain processes complex auditory information, like speech, is a major challenge. The cochlear nucleus (CN) is the first stage of auditory processing in the brain, and its neurons, particularly Globular Bushy Cells (GBCs), are known for their precise temporal coding. However, accurately quantifying the information these neurons carry, especially in response to dynamic stimuli like speech, is difficult. Traditional entropy estimation methods, like the "direct method," assume a memoryless source and are time-invariant, failing to capture the dynamic nature of neuronal responses and the temporal dependencies in speech signals. This paper aims to improve entropy estimation for neuronal spike trains to better reflect their information processing capabilities, particularly for GBCs responding to speech.

Methods

The authors employ a sophisticated computational model of the auditory system, specifically focusing on Globular Bushy Cells (GBCs) in the Cochlear Nucleus. This model simulates neuronal firing patterns in response to auditory inputs. The core of their contribution lies in developing an improved entropy estimation method.

Instead of relying on the standard "direct method" which provides a time-invariant upper bound, they introduce a time-varying and time-dependent entropy estimate. This approach accounts for:

  • Time-varying entropy: The entropy of neuronal firing can change over time, reflecting the changing nature of the auditory input.
  • Time-dependent entropy: Neuronal firing at a given moment is likely influenced by past firing patterns (i.e., the system has memory).

They achieve this by using a sliding window approach and applying the concept of conditional entropy. Specifically, they consider the entropy of a current "word" (a segment of the spike train) given past "words." This allows them to quantify the predictability of neuronal responses while explicitly accounting for temporal dependencies. They also investigate the impact of the window length (T) and the extent of conditioning (M) on the entropy estimate, which indirectly reveals the neuronal memory.

Datasets

The paper uses simulated spike trains generated from a model of the auditory system. The specific auditory inputs used for analysis are two utterances from the ISOLET spoken letter database:

  • The vowel sound /ay/ at a characteristic frequency (CF) of 1.5 kHz.
  • The fricative sound /es/ at a CF of 6.15 kHz.

The simulations allow for a large number of samples (on the order of 10^6) to be analyzed, which is crucial for accurate entropy estimation. No new datasets were introduced.

Key Insights and Ideas

  • High Entropy Rates Reflect Precise Coding: The GBC neurons exhibit remarkably high entropy rates in response to speech. This is not a sign of randomness but rather a reflection of their exquisite temporal precision in encoding complex sound features.
  • Time-Invariant Entropy is Insufficient: Standard, time-invariant entropy estimates fail to capture the dynamic nature of neuronal responses to speech. The proposed time-varying and time-dependent method provides a much richer picture.
  • Neuronal Memory Revealed: The improved entropy estimation method, particularly the conditional entropy, can reveal the temporal memory of the neuronal process. They demonstrate that GBCs exhibit correlations in their firing patterns up to at least 20 ms.
  • Vowel vs. Fricative Coding: The entropy estimates clearly differentiate between the coding of voiced sounds (like vowels, showing lower entropy due to regular firing patterns) and unvoiced sounds (like fricatives, exhibiting higher entropy due to their more noise-like components).
  • Predictability and Frequency Content: The analysis shows a strong correlation between the time-varying entropy estimate and the log-amplitude of the frequency content of the auditory signal, suggesting that entropy is a good proxy for how well the neuron's response predicts the input's characteristics.

Key Results

  • The time-varying and time-dependent entropy estimate shows a precise temporal correlation with the log-amplitude of the frequency content of the speech signal.
  • For voiced sounds like /ay/, the entropy estimate is lower due to regular firing patterns, while for noise-like fricatives like /s/, the entropy rate is higher.
  • The analysis of conditioning length (M) suggests that neuronal firings in GBCs are correlated up to at least 20 ms, indicating a significant neuronal memory.
  • The proposed method provides a better estimate of neuronal predictability compared to previous approaches, offering higher temporal resolution while still accounting for temporal dependencies.

Limitations and Future Work

  • Computational Complexity: While the proposed method improves accuracy, estimating conditional entropy for a large number of past words can be computationally intensive. The paper acknowledges the difficulty in precisely estimating entropy when considering the conditional entropy of the current word given all past words.
  • Model Limitations: The analysis is based on a simulated model of GBCs. While the model is validated against experimental data, future work could involve applying the improved entropy estimation to actual experimental neuronal recordings.
  • Future Work: The authors suggest that this improved entropy estimation method is particularly suitable for analyzing neuronal responses with high temporal resolution and long time dependencies. Future work could involve applying this to a wider range of auditory stimuli and neuronal types to further elucidate information processing in the auditory system.

Summary updated

BibTeX
@inproceedings{grigorescu2012improving,
  author = {Andrea Grigorescu and Marek Rudnicki and Michael Isik and Werner Hemmert and Stefano Rini},
  title = {{Improving the entropy estimate of neuronal firings of modeled cochlear nucleus neurons}},
  booktitle = {Interspeech},
  year = {2012},
}
Figure