🧬 biology

Direct dependencies between neurons explain activity

This study demonstrates that direct, non-interacting dependencies on individual inputs are sufficient to explain most neuronal activity variability across species and brain regions, suggesting that complex neural computations can be accurately modeled by simple logistic artificial neurons despite intricate biophysical details.

Original authors: Christopher W. Lynn

Published 2026-03-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Christopher W. Lynn

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine a neuron in your brain as a tiny, busy chef in a massive, chaotic kitchen.

For decades, scientists believed that this chef was incredibly complex. They thought the chef needed to taste every possible combination of ingredients at once to decide what to cook. If you added salt and pepper and garlic, the chef might react differently than if you added just salt and pepper. This idea suggested that the brain's "recipe book" was infinitely complicated, requiring the chef to remember millions of specific ingredient combinations.

The Big Discovery
Christopher Lynn's paper flips this script. He argues that the chef is actually much simpler. The chef doesn't need to worry about complex ingredient combinations. Instead, the chef just looks at each ingredient individually and decides: "Does salt make me want to cook? Yes. Does pepper? No."

The paper shows that if you know how a neuron reacts to each of its inputs (ingredients) one by one, you can predict almost everything it will do. You don't need to know how the inputs interact with each other.

The "Logistic Neuron" Analogy

Think of a neuron not as a biological mystery, but as a smart light switch.

The Inputs: Imagine 1,000 people standing around the switch, each holding a button.
The Old View: We thought the light would only turn on if a very specific, complex group of people pressed their buttons at the exact same time (e.g., "Only if Person A, B, and C press together, but not if D is there").
The New View: The paper shows that the light switch is actually just a simple calculator. It adds up the "votes" from each person. If the total "vote" crosses a certain line, the light turns on. It doesn't care about the group dynamics; it only cares about the sum of individual votes.

In math terms, this simple switch is called a logistic artificial neuron. The paper proves that real, biological neurons in mice, flies, and even worms act almost exactly like these simple artificial switches.

The "Noise" vs. "Signal" Metaphor

Imagine you are trying to understand a conversation in a crowded room.

The Complex View: You think you need to understand the complex slang, the inside jokes, and how everyone's voices blend together to understand what's being said.
The Paper's View: The paper says, "Actually, you can understand 90% of the conversation just by listening to who is speaking to whom, one-on-one."

The researchers found that direct dependencies (who talks to whom directly) explain about 90% of a neuron's activity. The remaining 10% (the complex interactions, the timing delays, and the hidden variables) is just "static" or noise. The brain is surprisingly efficient; it doesn't need a super-complex recipe to function.

Why This Matters: The "Redundant" Safety Net

Here is the most fascinating part: Because the system is so simple, it is also incredibly robust.

Imagine a bridge held up by 1,000 cables.

If the bridge relied on a complex, unique knot between every single cable, cutting just one cable might make the whole thing collapse.
But in this brain, the "cables" (inputs) are redundant. The paper found that you could cut away 90% of the connections to a neuron, and it would still work almost perfectly.

It's like a team of 100 people carrying a heavy piano. If the piano is carried by a complex, interlocking web of trust, losing one person is a disaster. But if they are just all pulling on the same rope, you can lose 10 people, and the piano keeps moving. The brain is built with a massive safety net, allowing it to survive damage or errors without failing.

The "XOR" Problem (The One Exception)

The paper uses a classic logic puzzle called XOR (Exclusive OR) to explain what they don't see.

AND Gate: "I will go out if it's Sunny AND I have money." (Simple, direct).
XOR Gate: "I will go out if it's Sunny OR I have money, but NOT both." (This requires a complex interaction between the two conditions).

The paper shows that real neurons are great at AND and OR (simple direct sums), but they are terrible at XOR. This confirms that neurons are mostly simple "adders" and not complex "logic processors" that need to understand intricate combinations.

The Bottom Line

Despite the fact that neurons have beautiful, intricate shapes and complex biology, their behavior is surprisingly simple.

The brain isn't a chaotic storm of complex interactions. It's more like a vast, efficient network of simple switches. By understanding how each neuron listens to its neighbors individually, we can predict almost everything the brain does. This suggests that the "code" of the brain is highly redundant and robust, designed to keep working even when parts of the network are damaged or noisy.

In short: The brain is less like a complex supercomputer trying to solve a million-variable equation, and more like a million simple light switches that, when turned on together, create the symphony of thought.

1. Problem Statement

The fundamental assumption in neural computation is that a neuron's output is determined by a linear summation of its inputs (the "integrate-and-fire" model). However, biological neurons possess complex dendritic structures and biophysical properties that theoretically allow for non-linear interactions between inputs (e.g., XOR-like logic, gating).

The central question addressed is: Do real neurons require complex interactions between inputs to explain their activity, or can their behavior be fully described by simple, direct dependencies on individual inputs?

While large-scale recordings now allow simultaneous observation of thousands of neurons, determining whether a neuron's activity arises from simple direct dependencies or complex higher-order interactions remains a challenge. Traditional models often assume interactions exist but lack a rigorous framework to quantify how much of the activity variability is actually explained by direct dependencies alone.

2. Methodology

The author develops a framework based on Maximum Entropy (MaxEnt) modeling to quantify the explanatory power of direct dependencies.

Binary Binarization: Neuronal activity is binarized into active ( $y=1$ ) or silent ( $y=0$ ) states within a time window $\Delta t$ .
The Minimal Model (MaxEnt): The study constructs a model that matches the measured direct dependencies $P(y|x_i)$ $P (y ∣ x_{i})$ (the probability of the output firing given a specific input $i$ $i$ ) but assumes maximum entropy (maximum randomness) regarding all higher-order interactions and time-delayed dependencies.
- Mathematically, this model takes the form of a logistic artificial neuron (equivalent to a perceptron):
  $P(y=1|x_1, \dots, x_n) = \sigma\left(b + \sum_{i=1}^n w_i x_i\right)$
  where $\sigma$ is the logistic function, $b$ is a bias, and $w_i$ are weights.
Entropy Hierarchy: The study defines a hierarchy of entropies to measure variability:
- $S_{tot}$ : Total entropy (uncertainty without knowing inputs).
- $S_{dir}$ : Entropy of the MaxEnt model constrained only by direct dependencies.
- $S_{true}$ : True entropy (latent variability not explainable by inputs).
- The difference $S_{tot} - S_{dir}$ represents the variability captured solely by direct dependencies.
Greedy Input Selection: Since synaptic connectivity is rarely known in large populations, the author uses a greedy algorithm to identify the minimal set of optimal inputs ( $n^*$ ) that minimizes $S_{dir}$ . The algorithm stops when the model can predict all other pairwise correlations in the population, preventing overfitting.
Datasets: The framework is applied to four distinct datasets:
1. Mouse Hippocampus (CA1, $N=1,485$ ).
2. Mouse Visual Cortex ( $N=11,445$ ) during natural image viewing.
3. Mouse Visual Cortex ( $N=11,445$ ) during spontaneous activity.
4. C. elegans whole brain ( $N=128$ ).

3. Key Contributions

Quantification of Direct Dependencies: The paper provides a rigorous mathematical metric ( $S_{tot} - S_{dir}$ ) to determine the fraction of neuronal variability explained by simple linear inputs versus complex interactions.
Equivalence to Artificial Neurons: It establishes that a biological neuron with purely direct dependencies is mathematically equivalent to a logistic artificial neuron (perceptron).
Emergence of Complexity: It demonstrates that complex higher-order correlations (triplets, quadruplets) and time-delayed dependencies are emergent properties of simple direct dependencies, rather than requiring intrinsic non-linear interactions.
Robustness and Redundancy: The inferred networks are shown to be highly redundant and robust to the removal of inputs (ablation), suggesting a sparse but fault-tolerant neural code.

4. Key Results

A. Direct Dependencies Explain Most Variability

Across all species and brain regions studied, direct dependencies explain the vast majority of neuronal activity:

Mouse Hippocampus: Direct dependencies explain >90% of the variability ( $S_{dir} \approx S_{true}$ ).
Mouse Visual Cortex: Similar results, with 91% of variability explained by direct dependencies for both natural images and spontaneous activity.
C. elegans: Even in this small, fully connected brain, 62% of variability is explained by direct dependencies using a minimal set of only 5 inputs per neuron.

B. Minimal Input Sets

The number of inputs required to capture all direct dependencies is surprisingly small relative to the total population:

Hippocampus: Median $n^* = 214$ inputs (14% of the population).
Visual Cortex: Median $n^* = 108$ inputs (<1% of the population).
C. elegans: Median $n^* = 5$ inputs.
Randomly selected inputs fail to explain this variability, confirming that the identified inputs are functionally relevant.

C. Prediction of Higher-Order and Time-Delayed Dependencies

The "complete" models (constrained only by direct dependencies) successfully predict complex statistics without explicitly modeling them:

Higher-Order Correlations: The models predict 99.85% of triplet correlations, 99.89% of quadruplet correlations, and 99.92% of quintuplet correlations in the hippocampus.
Time-Delayed Dependencies: Despite being constrained only to instantaneous dependencies, the models predict 99.6% of correlations at 0.1s delay, 99.1% at 1s, and 99.0% at 10s.
Conclusion: Complex temporal and combinatorial dynamics arise naturally from the superposition of simple, instantaneous direct dependencies.

D. Inferred Network Structure

The weights ( $w_i$ ) and biases ( $b$ ) inferred from the data exhibit four universal features of biological synaptic connectivity:

Sparsity: Only a small fraction of possible connections are significant.
Heavy-Tailed Distribution: Weight magnitudes follow a log-normal distribution (few strong, many weak).
Balance: Weights are evenly split between positive (excitatory) and negative (inhibitory).
Directionality: Connections are highly directed ( $w_{ij} \neq w_{ji}$ ).

E. Robustness

The inferred networks are strikingly robust. Simulating the ablation (removal) of up to 90% of a neuron's inputs results in minimal change to the predicted output activity, indicating a highly redundant neural code.

5. Significance

Simplification of Neural Computation: The findings suggest that despite the intricate biophysics of real neurons (dendrites, ion channels), their functional input-output mapping is effectively captured by simple linear models (perceptrons). This bridges the gap between biological neurons and simple artificial neural networks.
Implications for Neural Coding: The high redundancy and robustness imply that the brain does not rely on precise, unique combinations of inputs for computation but rather on a distributed, fault-tolerant code.
Methodological Advancement: The framework provides a tool to distinguish between "true" complex interactions and emergent correlations. It suggests that many phenomena previously attributed to complex non-linear dendritic processing may actually be explained by the statistical structure of the input population.
Future Directions: The paper opens avenues for investigating how these direct dependencies change with learning, how they relate to specific synaptic wiring maps (connectomics), and how they function in electrophysiological recordings with higher temporal resolution.

In summary, the paper argues that simple direct dependencies are the dominant driver of neuronal activity, and that the apparent complexity of neural computation is largely an emergent property of these simple linear interactions within a large, redundant network.