Artificial Neuron Architecture

Over the last two lessons, we engineered the two foundational stages of an artificial neuron:

The Linear Step (Lesson 1): We computed the dot product of features and weights ( $w \cdot x$ ), then added a baseline offset bias ( $+b$ ) to produce a raw linear score: $z = w \cdot x + b$ .
The Activation Function (Lesson 2): We passed that raw linear score through the Sigmoid function ( $\sigma(z) = \frac{1}{1 + e^{-z}}$ ) to transform it into an interpretable probability between $0$ and $1$ .

When you connect these two stages back-to-back, you get the complete mathematical anatomy of an Artificial Neuron.

The Unified Computational Pipeline

Every artificial neuron—from a single standalone classifier to the billions of units inside modern large language models—processes information through the exact same computational pipeline:

\text{Input } (x) \xrightarrow{\quad \text{Weights } (w) \quad} \sum_{i=1}^n w_i x_i \xrightarrow{\quad \text{Bias } (+b) \quad} z \xrightarrow{\quad \text{Activation } (\sigma) \quad} \text{Output } (\sigma(z))

Let's trace what happens at each stage of this journey:

Input Feature Vector ( $x \in \mathbb{R}^n$ ): The input to the neuron, containing the ordered feature values representing what we want to evaluate (e.g., our movie script specs $[1.0, 0.0, 0.5, 1.0]$ ).
Weight Vector ( $w \in \mathbb{R}^n$ ): The numerical measures of influence assigned to each feature.
Dot Product ( $w \cdot x = \sum w_i x_i$ ): Multiplies each feature value by its corresponding weight and adds them together into a single raw score.
Baseline Bias ( $b \in \mathbb{R}$ ): An independent baseline offset. It shifts the score up or down regardless of the input features.
Linear Score ( $z = w \cdot x + b$ ): The combined score after adding the baseline bias to the dot product.
Activation Function ( $\sigma$ ): A mathematical transformation applied to the linear score. In our classifier, we use the Sigmoid function to compress the unbounded score $z$ into a probability between $0$ and $1$ , while deeper networks use other activation functions (like ReLU) in this same slot.
Neuron Output ( $\sigma(z)$ ): The final scalar activation produced by the neuron after both steps.

(In our movie classifier, this continuous activation is compared against a threshold of $0.50$ to classify Hit vs Flop).

Tracing the End-to-End Computational Pipeline

To see how all these mathematical primitives connect into a single unified pipeline, trace how our two film scripts flow through feature weighting, dot product accumulation, baseline bias adjustment, Sigmoid compression, and threshold decision:

🎬 'Die Hard in Space': Complete Pipeline Trace

Feature	Feature Value (x)	Weight (w)	Product (w_i · x_i)	Dot Product (w · x)	Industry Bias (b)	Linear Score (z)	Sigmoid (σ(z))	Classification
Action	1.0	+5.0	+5.0	+9.5	−2.0	+7.5	0.999 (99.9%)	🟢 Hit (High Potential)
Romance	0.0	−2.0	0.0
Comedy	0.5	+1.0	+0.5
Sci-Fi	1.0	+4.0	+4.0

🎬 'The Notebook 2': Complete Pipeline Trace

Feature	Feature Value (x)	Weight (w)	Product (w_i · x_i)	Dot Product (w · x)	Industry Bias (b)	Linear Score (z)	Sigmoid (σ(z))	Classification
Action	0.0	+5.0	0.0	−1.5	−2.0	−3.5	0.029 (2.9%)	🔴 Flop (Low Potential)
Romance	1.0	−2.0	−2.0
Comedy	0.5	+1.0	+0.5
Sci-Fi	0.0	+4.0	0.0

Notice the clean division of labor across every column:

The Weights ( $w$ ) determine which features matter.
The Bias ( $b$ ) determines how hard it is to activate.
The Activation ( $\sigma$ ) determines the scale and range of the output.

NOTE: A Single Neuron Is Already a Complete AI Model

It takes an input vector, evaluates features, offsets a baseline, and classifies a prediction.

But a single neuron is not enough for complex real-world decisions. In Topics 3 and 4, we scale this building block by stacking multiple neurons together into layers and deep networks.