181 From Biological to Artificial Neurons

The artificial neuron sits at the conceptual root of every modern deep network, from convolutional vision models to transformer language models. Yet the object that practitioners compute with is a heavily idealized caricature of its biological namesake. Understanding both the inspiration and the distance between the two clarifies what artificial neural networks actually compute, why they are tractable, and where the biological analogy quietly breaks down. This chapter traces the lineage from the biological neuron to the McCulloch and Pitts threshold logic unit, then to the weighted sum plus nonlinearity that defines the modern artificial neuron, and finally examines the abstraction itself as a deliberate modeling choice with consequences.

181.1 1. The Biological Neuron, Briefly

A biological neuron is an electrically excitable cell specialized for receiving, integrating, and transmitting signals. Its functional anatomy can be summarized in four parts. The dendrites form a branching arbor that receives inputs from other neurons. The soma, or cell body, integrates those inputs. The axon carries an outgoing signal, often over considerable distance. The synapses are the junctions, typically chemical, where the axon terminal of one neuron influences the dendrite of another.

The signaling logic is worth stating with some care, because the artificial abstraction inherits part of it and discards the rest. Inputs arriving at synapses produce small, graded changes in the membrane potential of the receiving neuron, called postsynaptic potentials. These can be excitatory, nudging the membrane toward depolarization, or inhibitory, pulling it the other way. The soma sums these contributions across space and time. When the aggregated potential at the axon hillock crosses a threshold of roughly $-55$ millivolts, the neuron fires an action potential: a rapid, stereotyped, all-or-nothing electrical spike that propagates down the axon.

Three properties of this picture matter for what follows. First, integration is fundamentally additive and leaky: contributions accumulate but also decay. Second, the output is a threshold event, not a continuous quantity at the moment of firing. Third, information is carried not by the shape of any single spike, which is essentially fixed, but by the timing and rate of spikes in a train. A neuron that fires $80$ times per second is signaling something different from one firing $5$ times per second. The synapse itself is plastic: the efficacy with which a presynaptic spike influences the postsynaptic cell changes with experience, a phenomenon associated with Hebbian learning and long-term potentiation. This plasticity is the biological seed of the idea that adjustable weights can encode learning.

It is also worth being honest about complexity that the abstraction will ignore. Real neurons exhibit dendritic computation, where branches perform local nonlinear operations before signals reach the soma. They display refractory periods, adaptation, bursting, and a zoo of channel dynamics governed by the Hodgkin and Huxley equations. There are hundreds of morphologically and electrophysiologically distinct neuron types. None of this survives into the artificial neuron in any direct form.

181.2 2. The McCulloch and Pitts Neuron

In 1943, Warren McCulloch, a neurophysiologist, and Walter Pitts, a logician, published a paper proposing that the all-or-nothing firing of neurons could be treated as a logical proposition [1]. Their move was radical: if a neuron either fires or does not, then its activity is a Boolean variable, and a network of such neurons computes a logical function of its inputs. This collapsed the messy biophysics of section 1 into a discrete computational primitive.

181.2.1 2.1 Definition

A McCulloch and Pitts (MCP) unit takes binary inputs $x_1, x_2, \ldots, x_n \in \{0, 1\}$ and produces a binary output $y \in \{0, 1\}$. Each input is either excitatory or inhibitory. The unit has a fixed integer threshold $\theta$. The output is determined by comparing the sum of excitatory inputs against the threshold, with the crucial rule that any active inhibitory input vetoes firing absolutely.

In the simplest formulation with all inputs excitatory and unit weights, the rule is

\[ y = \begin{cases} 1 & \text{if } \displaystyle\sum_{i=1}^{n} x_i \geq \theta \\[4pt] 0 & \text{otherwise.} \end{cases} \]

The absolute inhibition rule reflected a then-current belief about how certain inhibitory synapses operate, and it gives the model a clean correspondence with logic.

181.2.2 2.2 Computing Logic

With this single template and a choice of $\theta$, the MCP unit realizes the basic logical connectives. For two excitatory inputs:

AND   : theta = 2   ->  fires only when x1 = x2 = 1
OR    : theta = 1   ->  fires when at least one input is 1
NOT   : one inhibitory input, theta = 0  ->  fires unless input is 1

The presence of NOT, together with AND and OR, makes the set functionally complete: any Boolean function can be assembled from these primitives. McCulloch and Pitts went further, showing that networks of such units, when augmented with cycles to provide memory, could in principle represent any computation expressible in their logical calculus. The conceptual payoff was immense. It suggested that the brain, viewed at the level of spikes, was an instance of a logical machine, and it placed neural modeling and the emerging theory of computation in the same frame, alongside the contemporaneous work of Turing.

181.2.3 2.3 What the MCP Unit Lacks

The MCP neuron is a fixed-function device. Its weights are implicitly $\pm 1$ and its threshold is set by hand to achieve a desired logic gate. There is no learning: nothing in the model adjusts itself in response to data. The synaptic plasticity that section 1 identified as biologically central is entirely absent. Inputs are strictly binary, and the absolute veto by inhibition is a brittle, all-or-nothing rule rather than a graded influence. These limitations are precisely what the next step in the lineage addresses.

181.3 3. The Artificial Neuron: Weighted Sum Plus Nonlinearity

The modern artificial neuron generalizes the MCP unit in two decisive ways. It replaces fixed unit connections with real-valued, adjustable weights, and it replaces the hard integer threshold with a general nonlinear activation function. The result is the unit used, with minor variations, throughout contemporary deep learning.

181.3.1 3.1 The Model

Given an input vector $\mathbf{x} = (x_1, \ldots, x_n)^\top \in \mathbb{R}^n$, the neuron is parameterized by a weight vector $\mathbf{w} = (w_1, \ldots, w_n)^\top \in \mathbb{R}^n$ and a scalar bias $b \in \mathbb{R}$. It first computes a pre-activation, the affine combination

\[ z = \sum_{i=1}^{n} w_i x_i + b = \mathbf{w}^\top \mathbf{x} + b, \]

and then applies a nonlinear activation function $\varphi : \mathbb{R} \to \mathbb{R}$ to produce the output

\[ a = \varphi(z) = \varphi\!\left(\mathbf{w}^\top \mathbf{x} + b\right). \]

The mapping from biology is now a loose analogy rather than a literal one. The weights $w_i$ play the role of synaptic efficacies, with positive weights excitatory and negative weights inhibitory. The weighted sum stands in for somatic integration. The bias $b$ shifts the effective threshold, since firing depends on whether $\mathbf{w}^\top \mathbf{x}$ exceeds $-b$. The activation $\varphi$ abstracts the firing decision.

inputs  x_i  ->  multiply by weights w_i  ->  sum + bias  =  z
z  ->  activation phi(z)  =  a   (the neuron's output)

181.3.2 3.2 The Role of the Bias and a Geometric View

It is convenient to absorb the bias into the weight vector by appending a constant input $x_0 = 1$ with weight $w_0 = b$, so that $z = \tilde{\mathbf{w}}^\top \tilde{\mathbf{x}}$. With this convention the neuron’s pre-activation is a single inner product.

Geometrically, the equation $\mathbf{w}^\top \mathbf{x} + b = 0$ defines a hyperplane in $\mathbb{R}^n$. The weight vector $\mathbf{w}$ is normal to this hyperplane, and the bias sets its offset from the origin. The sign of $z$ tells us which side of the hyperplane the input $\mathbf{x}$ lies on, and its magnitude is proportional to the signed distance. A single neuron with a threshold-like activation is therefore a linear classifier: it partitions input space into two half-spaces. This geometric reading recurs throughout machine learning and is the key to understanding both the power and the limits of a single unit.

181.3.3 3.3 Activation Functions

The choice of $\varphi$ determines what the neuron can express and how it can be trained. Several canonical choices form a historical and functional progression.

The Heaviside step function,

\[ \varphi(z) = \begin{cases} 1 & z \geq 0 \\ 0 & z < 0, \end{cases} \]

recovers the MCP-style threshold and yields Rosenblatt’s perceptron when paired with a learning rule. Its defect is that it is flat almost everywhere and discontinuous at the origin, so its derivative is zero or undefined. Gradient-based learning is impossible through it.

The logistic sigmoid,

\[ \sigma(z) = \frac{1}{1 + e^{-z}}, \]

is a smooth, differentiable surrogate that squashes its input into $(0, 1)$ and can be read as a firing probability. Its derivative $\sigma'(z) = \sigma(z)\bigl(1 - \sigma(z)\bigr)$ is convenient but saturates toward zero for large $|z|$, which throttles gradient flow in deep stacks. The hyperbolic tangent $\tanh(z)$ is a zero-centered relative with the same saturation issue.

The rectified linear unit,

\[ \mathrm{ReLU}(z) = \max(0, z), \]

has become the default in deep networks. It is cheap to compute, does not saturate for positive inputs, and its piecewise-linear form mitigates the vanishing-gradient problem that plagues sigmoidal units [4]. Variants such as leaky ReLU and GELU adjust its behavior near and below zero.

The essential point is that nonlinearity is not optional. If $\varphi$ were the identity, then a composition of neurons would collapse: stacking affine maps $\mathbf{W}_2(\mathbf{W}_1 \mathbf{x}) = (\mathbf{W}_2 \mathbf{W}_1)\mathbf{x}$ yields just another affine map. The expressive gain of depth comes entirely from inserting a nonlinearity between linear layers. The universal approximation results, which show that a single hidden layer of such units can approximate any continuous function on a compact domain to arbitrary accuracy, depend on $\varphi$ being nonconstant, bounded, and nonpolynomial [3].

181.3.4 3.4 From Unit to Network

A single neuron is a linear classifier and inherits a corresponding limitation, made famous by Minsky and Papert [2]: it cannot represent functions whose positive and negative examples are not linearly separable. The exclusive-or (XOR) function is the canonical counterexample, since no single hyperplane separates its two classes. The resolution is composition. Arranging neurons into layers, where the outputs of one layer feed the next, produces the multilayer perceptron, whose hidden units carve input space into regions that a final unit can combine. Learning then becomes the problem of choosing all the weights and biases jointly, which is accomplished by gradient descent on a loss function with gradients supplied by backpropagation. That machinery is the subject of later chapters; here it suffices that the artificial neuron, unlike the MCP unit, is trainable precisely because its weights are continuous and its activation is differentiable.

181.4 4. The Abstraction and Its Limits

The artificial neuron is a model, and like all models it is useful in proportion to what it deliberately ignores. It is worth tabulating the correspondence and, more importantly, the divergences.

181.4.1 4.1 What Survives the Abstraction

Three biological ideas carry through with real fidelity. Integration of weighted inputs survives: both the soma and the artificial unit sum scaled contributions. Excitation and inhibition survive as the signs of weights. Plasticity as the locus of learning survives, transformed from Hebbian synaptic modification into gradient-based weight updates. These three correspondences are enough to make the biological metaphor genuinely illuminating rather than merely decorative.

181.4.2 4.2 What Is Lost

The losses are substantial and should temper any claim that artificial networks are models of the brain. Real neurons communicate with spike trains in continuous time, encoding information in precise timing and rates; the standard artificial neuron emits a single static real number per forward pass and has no temporal dynamics at all. Biological signaling is stochastic and energy-constrained, with synaptic transmission failing probabilistically and the whole system operating on a power budget of roughly $20$ watts; the artificial unit is deterministic and, at scale, enormously power-hungry. Dendritic nonlinear computation, where a single biological neuron may itself implement something closer to a small multilayer network, is flattened into one scalar inner product. The brain’s connectivity is recurrent, sparse, and structured by development and experience, whereas the canonical artificial layer is dense and feedforward. And biological learning is local, using signals available at each synapse, while backpropagation requires a global backward pass that transmits error information through a path that has no clean biological counterpart. The biological plausibility of backpropagation remains an open research question.

181.4.3 4.3 Reading the Abstraction Correctly

The right way to hold this is that the artificial neuron was inspired by biology but is justified by mathematics and engineering, not by fidelity to the cell. Its value lies in being simultaneously expressive enough to compose into universal approximators and simple enough to differentiate and optimize at scale. The biological neuron motivated the form; the requirements of efficient learning fixed the details. Confusing the two leads to two opposite errors: dismissing artificial networks because they are unlike brains, and over-claiming that they explain cognition because they are called neural. The productive stance treats the neuron as what it is, a parameterized nonlinear function and a building block, while remaining curious about which of biology’s discarded features, such as spiking dynamics, local learning rules, and sparse recurrent structure, might yet be worth reclaiming. Neuromorphic computing and spiking neural networks pursue exactly that reclamation, and they remain a reminder that the abstraction of section 3 is a choice rather than a necessity.

181.5 5. Summary

The lineage is a sequence of deliberate simplifications followed by a re-enrichment. The biological neuron integrates graded, time-varying inputs and fires plastic, all-or-nothing spikes. The McCulloch and Pitts unit froze that into binary threshold logic, gaining computational clarity but losing learning. The artificial neuron restored adaptability by making weights continuous and the threshold a differentiable nonlinearity, recovering the weighted-sum-plus-nonlinearity form $a = \varphi(\mathbf{w}^\top \mathbf{x} + b)$ that underwrites modern deep learning. The cost of that tractability is a wide gap from biological reality, a gap that is a feature of the engineering rather than a defect to be apologized for, but one that every serious practitioner should keep in view.

181.6 References

McCulloch, W. S., and Pitts, W. (1943). A logical calculus of the ideas immanent in nervous activity. Bulletin of Mathematical Biophysics, 5, 115 to 133. https://link.springer.com/article/10.1007/BF02478259
Minsky, M., and Papert, S. (1969). Perceptrons: An Introduction to Computational Geometry. MIT Press. https://mitpress.mit.edu/9780262630221/perceptrons/
Hornik, K., Stinchcombe, M., and White, H. (1989). Multilayer feedforward networks are universal approximators. Neural Networks, 2(5), 359 to 366. https://www.sciencedirect.com/science/article/abs/pii/0893608089900208
Glorot, X., Bordes, A., and Bengio, Y. (2011). Deep sparse rectifier neural networks. Proceedings of the 14th International Conference on Artificial Intelligence and Statistics (AISTATS). https://proceedings.mlr.press/v15/glorot11a.html
Goodfellow, I., Bengio, Y., and Courville, A. (2016). Deep Learning. MIT Press. https://www.deeplearningbook.org/
Kandel, E. R., Schwartz, J. H., Jessell, T. M., Siegelbaum, S. A., and Hudspeth, A. J. (2013). Principles of Neural Science, 5th edition. McGraw-Hill. https://neurology.mhmedical.com/book.aspx?bookID=1049

# From Biological to Artificial Neurons The artificial neuron sits at the conceptual root of every modern deep network, from convolutional vision models to transformer language models. Yet the object that practitioners compute with is a heavily idealized caricature of its biological namesake. Understanding both the inspiration and the distance between the two clarifies what artificial neural networks actually compute, why they are tractable, and where the biological analogy quietly breaks down. This chapter traces the lineage from the biological neuron to the McCulloch and Pitts threshold logic unit, then to the weighted sum plus nonlinearity that defines the modern artificial neuron, and finally examines the abstraction itself as a deliberate modeling choice with consequences. ## 1. The Biological Neuron, Briefly A biological neuron is an electrically excitable cell specialized for receiving, integrating, and transmitting signals. Its functional anatomy can be summarized in four parts. The **dendrites** form a branching arbor that receives inputs from other neurons. The **soma**, or cell body, integrates those inputs. The **axon** carries an outgoing signal, often over considerable distance. The **synapses** are the junctions, typically chemical, where the axon terminal of one neuron influences the dendrite of another. The signaling logic is worth stating with some care, because the artificial abstraction inherits part of it and discards the rest. Inputs arriving at synapses produce small, graded changes in the membrane potential of the receiving neuron, called postsynaptic potentials. These can be excitatory, nudging the membrane toward depolarization, or inhibitory, pulling it the other way. The soma sums these contributions across space and time. When the aggregated potential at the axon hillock crosses a threshold of roughly $-55$ millivolts, the neuron fires an **action potential**: a rapid, stereotyped, all-or-nothing electrical spike that propagates down the axon. Three properties of this picture matter for what follows. First, integration is fundamentally additive and leaky: contributions accumulate but also decay. Second, the output is a threshold event, not a continuous quantity at the moment of firing. Third, information is carried not by the shape of any single spike, which is essentially fixed, but by the **timing and rate** of spikes in a train. A neuron that fires $80$ times per second is signaling something different from one firing $5$ times per second. The synapse itself is plastic: the efficacy with which a presynaptic spike influences the postsynaptic cell changes with experience, a phenomenon associated with Hebbian learning and long-term potentiation. This plasticity is the biological seed of the idea that adjustable weights can encode learning. It is also worth being honest about complexity that the abstraction will ignore. Real neurons exhibit dendritic computation, where branches perform local nonlinear operations before signals reach the soma. They display refractory periods, adaptation, bursting, and a zoo of channel dynamics governed by the Hodgkin and Huxley equations. There are hundreds of morphologically and electrophysiologically distinct neuron types. None of this survives into the artificial neuron in any direct form. ## 2. The McCulloch and Pitts Neuron In 1943, Warren McCulloch, a neurophysiologist, and Walter Pitts, a logician, published a paper proposing that the all-or-nothing firing of neurons could be treated as a logical proposition [1]. Their move was radical: if a neuron either fires or does not, then its activity is a Boolean variable, and a network of such neurons computes a logical function of its inputs. This collapsed the messy biophysics of section 1 into a discrete computational primitive. ### 2.1 Definition A McCulloch and Pitts (MCP) unit takes binary inputs $x_1, x_2, \ldots, x_n \in \{0, 1\}$ and produces a binary output $y \in \{0, 1\}$. Each input is either **excitatory** or **inhibitory**. The unit has a fixed integer threshold $\theta$. The output is determined by comparing the sum of excitatory inputs against the threshold, with the crucial rule that any active inhibitory input vetoes firing absolutely. In the simplest formulation with all inputs excitatory and unit weights, the rule is $$ y = \begin{cases} 1 & \text{if } \displaystyle\sum_{i=1}^{n} x_i \geq \theta \\[4pt] 0 & \text{otherwise.} \end{cases} $$ The absolute inhibition rule reflected a then-current belief about how certain inhibitory synapses operate, and it gives the model a clean correspondence with logic. ### 2.2 Computing Logic With this single template and a choice of $\theta$, the MCP unit realizes the basic logical connectives. For two excitatory inputs: ```text AND : theta = 2 -> fires only when x1 = x2 = 1 OR : theta = 1 -> fires when at least one input is 1 NOT : one inhibitory input, theta = 0 -> fires unless input is 1 ``` The presence of `NOT`, together with `AND` and `OR`, makes the set functionally complete: any Boolean function can be assembled from these primitives. McCulloch and Pitts went further, showing that networks of such units, when augmented with cycles to provide memory, could in principle represent any computation expressible in their logical calculus. The conceptual payoff was immense. It suggested that the brain, viewed at the level of spikes, was an instance of a logical machine, and it placed neural modeling and the emerging theory of computation in the same frame, alongside the contemporaneous work of Turing. ### 2.3 What the MCP Unit Lacks The MCP neuron is a fixed-function device. Its weights are implicitly $\pm 1$ and its threshold is set by hand to achieve a desired logic gate. There is **no learning**: nothing in the model adjusts itself in response to data. The synaptic plasticity that section 1 identified as biologically central is entirely absent. Inputs are strictly binary, and the absolute veto by inhibition is a brittle, all-or-nothing rule rather than a graded influence. These limitations are precisely what the next step in the lineage addresses. ## 3. The Artificial Neuron: Weighted Sum Plus Nonlinearity The modern artificial neuron generalizes the MCP unit in two decisive ways. It replaces fixed unit connections with **real-valued, adjustable weights**, and it replaces the hard integer threshold with a general **nonlinear activation function**. The result is the unit used, with minor variations, throughout contemporary deep learning. ### 3.1 The Model Given an input vector $\mathbf{x} = (x_1, \ldots, x_n)^\top \in \mathbb{R}^n$, the neuron is parameterized by a weight vector $\mathbf{w} = (w_1, \ldots, w_n)^\top \in \mathbb{R}^n$ and a scalar **bias** $b \in \mathbb{R}$. It first computes a **pre-activation**, the affine combination $$ z = \sum_{i=1}^{n} w_i x_i + b = \mathbf{w}^\top \mathbf{x} + b, $$ and then applies a nonlinear activation function $\varphi : \mathbb{R} \to \mathbb{R}$ to produce the output $$ a = \varphi(z) = \varphi\!\left(\mathbf{w}^\top \mathbf{x} + b\right). $$ The mapping from biology is now a loose analogy rather than a literal one. The weights $w_i$ play the role of synaptic efficacies, with positive weights excitatory and negative weights inhibitory. The weighted sum stands in for somatic integration. The bias $b$ shifts the effective threshold, since firing depends on whether $\mathbf{w}^\top \mathbf{x}$ exceeds $-b$. The activation $\varphi$ abstracts the firing decision. ```text inputs x_i -> multiply by weights w_i -> sum + bias = z z -> activation phi(z) = a (the neuron's output) ``` ### 3.2 The Role of the Bias and a Geometric View It is convenient to absorb the bias into the weight vector by appending a constant input $x_0 = 1$ with weight $w_0 = b$, so that $z = \tilde{\mathbf{w}}^\top \tilde{\mathbf{x}}$. With this convention the neuron's pre-activation is a single inner product. Geometrically, the equation $\mathbf{w}^\top \mathbf{x} + b = 0$ defines a **hyperplane** in $\mathbb{R}^n$. The weight vector $\mathbf{w}$ is normal to this hyperplane, and the bias sets its offset from the origin. The sign of $z$ tells us which side of the hyperplane the input $\mathbf{x}$ lies on, and its magnitude is proportional to the signed distance. A single neuron with a threshold-like activation is therefore a **linear classifier**: it partitions input space into two half-spaces. This geometric reading recurs throughout machine learning and is the key to understanding both the power and the limits of a single unit. ### 3.3 Activation Functions The choice of $\varphi$ determines what the neuron can express and how it can be trained. Several canonical choices form a historical and functional progression. The **Heaviside step** function, $$ \varphi(z) = \begin{cases} 1 & z \geq 0 \\ 0 & z < 0, \end{cases} $$ recovers the MCP-style threshold and yields Rosenblatt's perceptron when paired with a learning rule. Its defect is that it is flat almost everywhere and discontinuous at the origin, so its derivative is zero or undefined. Gradient-based learning is impossible through it. The **logistic sigmoid**, $$ \sigma(z) = \frac{1}{1 + e^{-z}}, $$ is a smooth, differentiable surrogate that squashes its input into $(0, 1)$ and can be read as a firing probability. Its derivative $\sigma'(z) = \sigma(z)\bigl(1 - \sigma(z)\bigr)$ is convenient but saturates toward zero for large $|z|$, which throttles gradient flow in deep stacks. The hyperbolic tangent $\tanh(z)$ is a zero-centered relative with the same saturation issue. The **rectified linear unit**, $$ \mathrm{ReLU}(z) = \max(0, z), $$ has become the default in deep networks. It is cheap to compute, does not saturate for positive inputs, and its piecewise-linear form mitigates the vanishing-gradient problem that plagues sigmoidal units [4]. Variants such as leaky ReLU and GELU adjust its behavior near and below zero. The essential point is that **nonlinearity is not optional**. If $\varphi$ were the identity, then a composition of neurons would collapse: stacking affine maps $\mathbf{W}_2(\mathbf{W}_1 \mathbf{x}) = (\mathbf{W}_2 \mathbf{W}_1)\mathbf{x}$ yields just another affine map. The expressive gain of depth comes entirely from inserting a nonlinearity between linear layers. The universal approximation results, which show that a single hidden layer of such units can approximate any continuous function on a compact domain to arbitrary accuracy, depend on $\varphi$ being nonconstant, bounded, and nonpolynomial [3]. ### 3.4 From Unit to Network A single neuron is a linear classifier and inherits a corresponding limitation, made famous by Minsky and Papert [2]: it cannot represent functions whose positive and negative examples are not linearly separable. The exclusive-or (XOR) function is the canonical counterexample, since no single hyperplane separates its two classes. The resolution is composition. Arranging neurons into layers, where the outputs of one layer feed the next, produces the multilayer perceptron, whose hidden units carve input space into regions that a final unit can combine. Learning then becomes the problem of choosing all the weights and biases jointly, which is accomplished by gradient descent on a loss function with gradients supplied by backpropagation. That machinery is the subject of later chapters; here it suffices that the artificial neuron, unlike the MCP unit, is **trainable** precisely because its weights are continuous and its activation is differentiable. ## 4. The Abstraction and Its Limits The artificial neuron is a model, and like all models it is useful in proportion to what it deliberately ignores. It is worth tabulating the correspondence and, more importantly, the divergences. ### 4.1 What Survives the Abstraction Three biological ideas carry through with real fidelity. **Integration of weighted inputs** survives: both the soma and the artificial unit sum scaled contributions. **Excitation and inhibition** survive as the signs of weights. **Plasticity as the locus of learning** survives, transformed from Hebbian synaptic modification into gradient-based weight updates. These three correspondences are enough to make the biological metaphor genuinely illuminating rather than merely decorative. ### 4.2 What Is Lost The losses are substantial and should temper any claim that artificial networks are models of the brain. Real neurons communicate with **spike trains in continuous time**, encoding information in precise timing and rates; the standard artificial neuron emits a single static real number per forward pass and has no temporal dynamics at all. Biological signaling is **stochastic and energy-constrained**, with synaptic transmission failing probabilistically and the whole system operating on a power budget of roughly $20$ watts; the artificial unit is deterministic and, at scale, enormously power-hungry. **Dendritic nonlinear computation**, where a single biological neuron may itself implement something closer to a small multilayer network, is flattened into one scalar inner product. The brain's connectivity is **recurrent, sparse, and structured** by development and experience, whereas the canonical artificial layer is dense and feedforward. And biological learning is **local**, using signals available at each synapse, while backpropagation requires a global backward pass that transmits error information through a path that has no clean biological counterpart. The biological plausibility of backpropagation remains an open research question. ### 4.3 Reading the Abstraction Correctly The right way to hold this is that the artificial neuron was inspired by biology but is justified by mathematics and engineering, not by fidelity to the cell. Its value lies in being simultaneously expressive enough to compose into universal approximators and simple enough to differentiate and optimize at scale. The biological neuron motivated the form; the requirements of efficient learning fixed the details. Confusing the two leads to two opposite errors: dismissing artificial networks because they are unlike brains, and over-claiming that they explain cognition because they are called neural. The productive stance treats the neuron as what it is, a parameterized nonlinear function and a building block, while remaining curious about which of biology's discarded features, such as spiking dynamics, local learning rules, and sparse recurrent structure, might yet be worth reclaiming. Neuromorphic computing and spiking neural networks pursue exactly that reclamation, and they remain a reminder that the abstraction of section 3 is a choice rather than a necessity. ## 5. Summary The lineage is a sequence of deliberate simplifications followed by a re-enrichment. The biological neuron integrates graded, time-varying inputs and fires plastic, all-or-nothing spikes. The McCulloch and Pitts unit froze that into binary threshold logic, gaining computational clarity but losing learning. The artificial neuron restored adaptability by making weights continuous and the threshold a differentiable nonlinearity, recovering the weighted-sum-plus-nonlinearity form $a = \varphi(\mathbf{w}^\top \mathbf{x} + b)$ that underwrites modern deep learning. The cost of that tractability is a wide gap from biological reality, a gap that is a feature of the engineering rather than a defect to be apologized for, but one that every serious practitioner should keep in view. ## References 1. McCulloch, W. S., and Pitts, W. (1943). A logical calculus of the ideas immanent in nervous activity. Bulletin of Mathematical Biophysics, 5, 115 to 133. https://link.springer.com/article/10.1007/BF02478259 2. Minsky, M., and Papert, S. (1969). Perceptrons: An Introduction to Computational Geometry. MIT Press. https://mitpress.mit.edu/9780262630221/perceptrons/ 3. Hornik, K., Stinchcombe, M., and White, H. (1989). Multilayer feedforward networks are universal approximators. Neural Networks, 2(5), 359 to 366. https://www.sciencedirect.com/science/article/abs/pii/0893608089900208 4. Glorot, X., Bordes, A., and Bengio, Y. (2011). Deep sparse rectifier neural networks. Proceedings of the 14th International Conference on Artificial Intelligence and Statistics (AISTATS). https://proceedings.mlr.press/v15/glorot11a.html 5. Goodfellow, I., Bengio, Y., and Courville, A. (2016). Deep Learning. MIT Press. https://www.deeplearningbook.org/ 6. Kandel, E. R., Schwartz, J. H., Jessell, T. M., Siegelbaum, S. A., and Hudspeth, A. J. (2013). Principles of Neural Science, 5th edition. McGraw-Hill. https://neurology.mhmedical.com/book.aspx?bookID=1049