Synchronisation, binding and the role of correlated firing in fast information transmission

Simon R. Schultz, Huw Golledge, Stefano Panzeri

Abstract: Does synchronization between action potentials from different neurons in the visual system play a substantial role in solving the binding problem? The binding problem can be studied quantitatively in the broader framework of the information contained in neural spike trains about some external correlate, which in this case is object configurations in the visual field. We approach this problem by using a mathematical formalism that quantifies the impact of correlated firing in short time scales. Using a power series expansion, the mutual information an ensemble of neurons conveys about external stimuli is broken down into firing rate and correlation components. This leads to a new quantification procedure directly applicable to simultaneous multiple neuron recordings. It theoretically constrains the neural code, showing that correlations contribute less significantly than firing rates to rapid information processing. By using this approach to study the limits upon the amount of information that an ideal observer is able to extract from a synchrony code, it may be possible to determine whether the available amount of information is sufficient to support computational processes such as feature binding.

The Binding Problem and Neural Synchrony

One of the most enduring mysteries in neuroscience is the "binding problem." When an observer perceives a complex scene, such as a red ball bouncing, the brain processes distinct features like color, shape, and motion in specialized, anatomically separate areas. The fundamental question is how the brain "binds" these disparate features together to form a coherent perception of a single object.

Illustration of the binding problem: a visual scene $S_0$ contains objects $S_1$ and $S_2$. The brain must correctly associate features (represented by dashed ellipses) with the correct object.

The "binding-by-synchrony" hypothesis proposes a temporal solution: neurons representing features of the same object synchronize their action potentials. While experimental evidence has shown instances of such synchrony, quantifying its actual contribution to information transmission has remained challenging. Schultz, Golledge, and Panzeri (2000) address this by applying information theory to determine if neural synchrony is a viable mechanism for rapid information processing.

Mathematical Framework: Partitioning Information

The authors' core contribution is a mathematical formalism that expands Mutual Information ($I$) as a power series in time ($t$):

$$I(t) = t I_t + \frac{1}{2} t^2 I_{tt} + \dots$$

This approach allows researchers to take simultaneous recordings from multiple neurons and explicitly partition the information into distinct components:

  • The First Order ($I_t$): This represents the "rate component." It shows that at the very start of a stimulus, the information rate depends entirely on the individual firing rates of the neurons.
  • The Second Order ($I_{tt}$): This contains the "correlational components." It accounts for both "signal correlations" (how mean rates across stimuli are related) and "noise correlations" (how trial-to-trial variability is shared between neurons).

Analysis of Information Transmission

The paper primarily uses simulated data, including Integrate-and-Fire (I&F) models of neuron quintuplets, to validate the framework. They tested scenarios ranging from standard shared input (30%) to "synchrony-heavy" cases where firing rates were kept constant but shared input reached 90%.

Dominance of Firing Rates

In short time windows (20–50 ms)—the timescales critical for rapid visual perception—the "rate" component was found to dominate. Even in the "best-case" scenario for synchrony (where only correlations changed and rates were fixed), the total amount of information conveyed was modest compared to what a standard rate-modulated code could achieve at the same firing frequency.

The Ideal Observer Limit

The information calculated here represents an upper bound. It represents the maximum amount of information any decoder could possibly extract. If this value is low for synchrony, then no biological mechanism—no matter how specialized—can extract more. The results suggest that for a correlational code to be truly efficient, the system would need to be strongly anti-correlated, which is rarely observed in the cortex.

Conclusion

By quantifying the information content of neural signals, this research theoretically constrains the neural code. It demonstrates that firing rates carry the vast majority of information during rapid processing, suggesting that the "binding-by-synchrony" hypothesis may face significant information-theoretic hurdles as a primary mechanism for fast visual binding. While synchrony may play a role in longer integration times, it appears to be a "second-order" effect in the rapid timescales required for perception.

Summary updated

Abstract

Does synchronization between action potentials from different neurons in the visual system play a substantial role in solving the binding problem? The binding problem can be studied quantitatively in the broader framework of the information contained in neural spike trains about some external correlate, which in this case is object configurations in the visual field. We approach this problem by using a mathematical formalism that quantifies the impact of correlated firing in short time scales. Using a power series expansion, the mutual information an ensemble of neurons conveys about external stimuli is broken down into firing rate and correlation components. This leads to a new quantification procedure directly applicable to simultaneous multiple neuron recordings. It theoretically constrains the neural code, showing that correlations contribute less significantly than firing rates to rapid information processing. By using this approach to study the limits upon the amount of information that an ideal observer is able to extract from a synchrony code, it may be possible to determine whether the available amount of information is sufficient to support computational processes such as feature binding.
BibTeX
@inproceedings{schultz2001synchronisation,
  author = {Simon R. Schultz and Huw Golledge and Stefano Panzeri},
  title = {{Synchronisation, binding and the role of correlated firing in fast information transmission}},
  booktitle = {Emergent Neural Computational Architectures Based on Neuroscience},
  year = {2001},
  doi = {10.1007/3-540-44597-8\_16},
}
Figure