Correlational Strength and Computational Algebra of Synaptic Connections Between Neurons

Eberhard E. FetzNeurIPS 1987

Overview

This paper establishes a "computational algebra" for neural networks, demonstrating that the statistical effect of sequential synaptic connections can be calculated via convolution, while parallel pathways combine through simple addition. It reveals a striking biological constraint: individual synapses are surprisingly weak, triggering a spike only about 1% of the time. This implies that reliable biological signal propagation requires massive parallelism.

Bridging the Microscopic and Macroscopic

In neuroscience, researchers often observe the brain at two distinct scales: the microscopic physical shape of a single voltage change in a neuron (an Excitatory Postsynaptic Potential, or EPSP) and the macroscopic statistical correlation between the firing of two neurons. The critical gap between these scales is the "transfer function"—how exactly does the physical shape of a synaptic potential translate into the probability of a neuron firing a spike?

Relationship between EPSP and cross-correlogram

Eberhard Fetz tackles this by investigating the relationship between EPSPs and cross-correlograms (histograms showing how much more likely a "downstream" neuron is to fire after an "upstream" neuron spikes). He seeks to define the rules for how these individual connections combine into complex circuits like disynaptic links, feedback loops, and parallel pathways.

Methodology and Data

As a foundational neurophysiology paper from 1988, the study relies on empirical biological data rather than modern large-scale benchmarks. The methodology synthesizes several experimental and theoretical approaches:

  • Intracellular Recording: Measuring actual voltage changes (EPSPs and IPSPs) in cat spinal cord motoneurons and neocortical pyramidal cells.
  • Cross-Correlation Analysis: Using extracellular spike trains to measure the "correlational strength" between connected neurons.
  • Mathematical Modeling & Simulation: Developing a threshold-crossing model where a neuron fires when its membrane potential plus the EPSP hits a specific limit. By adding "synaptic noise" (random background fluctuations) to the model, the simulations successfully broadened the predicted correlation peaks to match real-world biological data.

Key Insights: The Algebra of Connectivity

The paper introduces several foundational concepts for understanding neural circuitry:

  • The "Derivative Rule": In a noise-free system, the increase in firing probability is proportional to the derivative (the rate of rise) of the EPSP, rather than just its peak height. While synaptic noise "smears" this peak in reality, the area of the peak (the total number of triggered spikes) remains proportional to the EPSP's amplitude.
  • The Algebra of Connectivity: Fetz proposes a simple set of rules for network analysis. For sequential (series) connections, the correlation of a multi-step path (e.g., A → C → B) is the convolution of the individual links. For parallel connections, the effects of multiple pathways between the same two neurons are additive.
Monosynaptic vs. Disynaptic connections and their resulting correlations
  • The 1% Rule: A typical "unitary" EPSP (from a single input fiber) is about 100 $\mu$V. This only increases the firing probability by about 0.01 (1%). This is a crucial constraint for neural modeling, highlighting that single synapses are remarkably "unreliable" on their own.
  • Symmetry in Time: The probability of neuron B firing after A is identical to the time-inverted probability of A firing before B. This symmetry allows the same convolution math to work regardless of the direction of impulse propagation.

Results and Implications

The study found a linear scaling relationship where the number of triggered spikes ($N_p$) is linearly related to EPSP amplitude ($h$): $N_p \approx (0.1/mv) \cdot h(mv) + .003$.

Because a single disynaptic link (A → C → B) has a combined probability of only $10^{-4}$ ($0.01 \times 0.01$), Fetz concludes that "secure" propagation of activity through the brain strictly requires hundreds of parallel neurons at every level to compensate for the weakness of individual synapses. Furthermore, while positive feedback loops can be modeled as successive convolutions, they can only sustain activity if the number of interacting cells is large enough to overcome this low per-synapse firing probability.

Limitations and Future Directions

  • The Inverse Problem: While we can predict the correlation if we know the circuitry (the "forward problem"), deducing the exact circuitry from correlation data alone is much harder. Different circuit topologies can produce nearly identical correlation peaks.
  • Inhibition: While Inhibitory Postsynaptic Potentials (IPSPs) produce troughs in correlation, these effects had not yet been analyzed as quantitatively as excitatory ones at the time of publication.
  • Modulation: The proposed algebra assumes a certain baseline firing rate ($f_0$). If the background activity of the network changes, the "gain" of these synaptic links also changes, suggesting that the computational algebra of the brain is highly dynamic rather than static.

Summary updated

BibTeX
@inproceedings{fetz1987correlational,
  author = {Eberhard E. Fetz},
  title = {{Correlational Strength and Computational Algebra of Synaptic Connections Between Neurons}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {1987},
}
Figure