Synaptic Transmission: An Information-Theoretic Perspective

Amit Manwani, Christof KochNeurIPS 1997

Traditionally, neuroscience has often treated neurons as "black boxes," focusing primarily on the relationship between input and output spikes. This abstraction, however, overlooks the complex biophysical hardware of the brain—specifically, the synapse.

Cortical synapses are notoriously unreliable. Unlike the robust junctions in our muscles, a single electrical pulse in a brain cell often fails to release a neurotransmitter vesicle. Even when a release is successful, the size of the resulting electrical signal (the Excitatory Post-Synaptic Potential, or EPSP) varies wildly. By treating the synapse as a noisy communication channel, this paper applies information theory to quantify the reliability of cortical synapses, asking a fundamental question: How much information can such an unreliable component actually transmit?

Modeling the Synapse as a Communication Channel

To bridge the gap between biophysics and information theory, the authors construct a mathematical model of a cortical synapse consisting of three distinct stages:

  1. A Binary Channel: This models the "all-or-nothing" nature of vesicle release, accounting for both the probability of transmission failure and spontaneous, unprompted release.
  2. A Random Amplitude Filter: This captures the trial-to-trial variability in the size of the post-synaptic response, modeled using a Gamma distribution.
  3. Additive Gaussian Noise: This represents the background thermal and electrical noise present at the post-synaptic membrane.

The model is evaluated under two distinct coding paradigms. The first is Signal Estimation, where a continuous signal is encoded in the firing rate of a Poisson neuron, and the goal is to reconstruct the original signal using an optimal linear estimator (a Wiener Filter). The second is Signal Detection, a simpler "Yes/No" task where the system uses an optimal spike detector (a Matched Filter) to decide if a single presynaptic spike occurred.

Block diagram of the synaptic communication channel model

Rather than relying on a new large-scale dataset, this theoretical model is grounded in empirically measured parameter values from established literature. For instance, vesicle release probabilities typically range between 0.1 and 0.4, combined with established coefficients of variation for EPSP amplitudes and membrane time constants.

Key Insights: The Power of Redundancy

The study reveals that signal transmission is primarily limited by "shot noise"—the inherent randomness of discrete spikes and vesicle releases. Synaptic variability acts as a multiplier on this noise, meaning a single synapse operates in a fundamentally "low Signal-to-Noise Ratio" (SNR) regime.

In this state, the information rate is remarkably low. The authors derived closed-form expressions for the lower bounds of information capacity, finding that a single synapse often transmits at less than 5 bits/sec. Furthermore, increasing the signal's bandwidth does little to improve this rate.

Information rate and normalized error as a function of firing rate

However, the most striking insight is how quickly performance scales with redundancy. While a single synapse is a poor transmitter, adding just a small amount of redundancy—such as five synapses acting in parallel—moves the system into a "high-SNR" regime. Moving from $n=1$ to $n=5$ synapses yields a substantial jump in the mutual information rate (e.g., from ~2 bits/sec to ~15 bits/sec in some configurations). Once in this regime, information capacity begins to increase linearly with the number of synapses.

Limitations and Future Directions

The information rates calculated for these cortical models are significantly lower than those measured in peripheral sensory neurons (like those in the eye or ear). This efficiency gap suggests that the cortex relies heavily on redundancy and population coding rather than the high-fidelity transmission of individual connections.

While this model provides a powerful framework, the authors note a few limitations:

  • Linearity Constraints: The study focuses on optimal linear estimators. It is entirely possible that non-linear decoding mechanisms in the brain could extract even more information.
  • Simplified Dynamics: The model treats synaptic parameters as constants, ignoring history-dependent effects like paired-pulse facilitation (strengthening with rapid use) or vesicle depletion (weakening).
  • Conservative Estimates: The chosen parameters represent a "worst-case" scenario; actual biological synapses might be slightly more efficient.

Ultimately, this research lays the groundwork for a more comprehensive biophysical map of information flow, suggesting future extensions into more complex dendritic structures and dynamic spiking mechanisms.

Summary updated

BibTeX
@inproceedings{manwani1997synaptic,
  author = {Amit Manwani and Christof Koch},
  title = {{Synaptic Transmission: An Information-Theoretic Perspective}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {1997},
}
Figure