Matrix Dimensions and Parallel Vectors

In Topic 2, we built a complete single artificial neuron. We multiplied an input feature vector $x$ by a weight vector $w$ , added an independent baseline bias $b$ , and passed the resulting linear score $z$ through an activation function to yield a single scalar output.

In neural networks, this output is denoted by $a$ (short for activation):

z = w \cdot x + b \qquad \longrightarrow \qquad a = \sigma(z)

A single artificial neuron outputs a single scalar value. Because it computes only one dot product, it can only answer one specific question about an input. In Topic 2, that neuron answered: "Will 'Die Hard in Space' be a hit?"

But for most real-world tasks, a single output is not enough. What if Netflix wants to evaluate that same script from multiple angles? It might ask three distinct questions:

Will this movie be a hit?
Does this script have Oscar potential?
Can this concept launch a multi-film franchise?

Because each question values the script's features differently—audiences reward Action, while critics reward Romance—each outcome requires its own dedicated weight vector. To evaluate three separate questions, we need three separate neurons.

Evaluating these neurons one by one using separate, disconnected equations is slow and cumbersome. To compute all three predictions in parallel, we stack each neuron's weight vector as a horizontal row into a two-dimensional grid called a Matrix.

When it comes to AI, a matrix is a 2D rectangular grid of numbers organized into horizontal rows and vertical columns. Stacking multiple neuron weight vectors into a matrix allows a neural network to evaluate multiple outcomes on the exact same input features simultaneously.

Track 1 (Underlying Mechanism): The 3-Supermarket Grocery Bill

To see why stacking vectors into a matrix is useful, let's return to our shopping cart example from Topic 1.

Imagine your weekly grocery shopping profile is captured by the 3-dimensional feature vector $x \in \mathbb{R}^3$ :

x = \begin{bmatrix} 2 \\ 3 \\ 1 \end{bmatrix} \begin{matrix} \text{(Apples)} \\ \text{(Bananas)} \\ \text{(Carrots)} \end{matrix}

In Topic 1, we calculated your total bill at a single store by taking the dot product of your cart vector $x$ and that store's unit price vector $w$ .

Now, suppose you want to price-shop across three different grocery stores (Target, Walmart, and Whole Foods) at the same time to determine where your cart is cheapest. Each store sets its own independent unit prices for the three items:

Target Price Vector ( $w_1$ ): $w_1 = \begin{bmatrix} 4 \\ 1 \\ 2 \end{bmatrix} \implies w_1 \cdot x = (4 \times 2) + (1 \times 3) + (2 \times 1) = 8 + 3 + 2 = \char36 13$
Walmart Price Vector ( $w_2$ ): $w_2 = \begin{bmatrix} 3 \\ 1 \\ 1 \end{bmatrix} \implies w_2 \cdot x = (3 \times 2) + (1 \times 3) + (1 \times 1) = 6 + 3 + 1 = \char36 10$
Whole Foods Price Vector ( $w_3$ ): $w_3 = \begin{bmatrix} 6 \\ 2 \\ 4 \end{bmatrix} \implies w_3 \cdot x = (6 \times 2) + (2 \times 3) + (4 \times 1) = 12 + 6 + 4 = \char36 22$

Instead of computing three separate, disconnected dot product equations, we arrange each store's price vector as a horizontal row. Stacking these three rows on top of each other creates a single Price Matrix ( $W_{\text{grocery}} \in \mathbb{R}^{3 \times 3}$ ):

W_{\text{grocery}} = \begin{bmatrix} 4 & 1 & 2 \\ 3 & 1 & 1 \\ 6 & 2 & 4 \end{bmatrix} \begin{matrix} \text{(Row 1: Target Prices)} \\ \text{(Row 2: Walmart Prices)} \\ \text{(Row 3: Whole Foods Prices)} \end{matrix}

When we multiply this Price Matrix by your single shopping cart vector $x$ , we execute all three dot products in a single operation, producing an ordered vector of checkout totals:

W_{\text{grocery}} x = \begin{bmatrix} 4 & 1 & 2 \\ 3 & 1 & 1 \\ 6 & 2 & 4 \end{bmatrix} \begin{bmatrix} 2 \\ 3 \\ 1 \end{bmatrix} = \begin{bmatrix} (4 \times 2) + (1 \times 3) + (2 \times 1) \\ (3 \times 2) + (1 \times 3) + (1 \times 1) \\ (6 \times 2) + (2 \times 3) + (4 \times 1) \end{bmatrix} = \begin{bmatrix} 13 \\ 10 \\ 22 \end{bmatrix} \begin{matrix} \text{(\char36 13 at Target)} \\ \text{(\char36 10 at Walmart)} \\ \text{(\char36 22 at Whole Foods)} \end{matrix}

Notice the structure:

The shopping cart vector $x$ holds the feature values (the concrete quantities of each item).
The matrix $W_{\text{grocery}}$ holds the unit prices across all three stores, where each row represents one store's prices.
The output is an ordered vector representing the three simultaneous checkout totals.

Track 2 (Applied Concept): The 3 Studio Predictions

Let's port this mechanism directly into our running Netflix Greenlight Predictor.

In Topics 1 and 2, our input feature vector $x \in \mathbb{R}^4$ represented a script's feature values across four permanently locked genre dimensions [Action, Romance, Comedy, Sci-Fi]:

x_{\text{Die Hard}} = \begin{bmatrix} 1.0 \\ 0.0 \\ 0.5 \\ 1.0 \end{bmatrix} \begin{matrix} \text{(Action)} \\ \text{(Romance)} \\ \text{(Comedy)} \\ \text{(Sci-Fi)} \end{matrix}

Instead of calculating a single score, the studio executive committee evaluates three distinct commercial outcomes for every script simultaneously:

Hit ( $w_1$ ): Mainstream box office appeal. High Action and Sci-Fi boost the score; heavy Romance penalizes it. $w_1 = \begin{bmatrix} 5.0 \\ -2.0 \\ 1.0 \\ 4.0 \end{bmatrix}$
Oscar ( $w_2$ ): Critical acclaim and Academy Award prestige. High Romance and Comedy are rewarded; explosive Action is heavily penalized. $w_2 = \begin{bmatrix} -5.0 \\ 5.0 \\ 5.0 \\ 0.0 \end{bmatrix}$
Franchise ( $w_3$ ): Spinoff, sequel, and merchandising potential. Driven exclusively by high Action and high Sci-Fi worldbuilding. $w_3 = \begin{bmatrix} 5.0 \\ 0.0 \\ 0.0 \\ 4.0 \end{bmatrix}$

Just as we stacked the three grocery store price vectors into rows, we arrange these three studio weight vectors horizontally as rows into a single Weight Matrix ( $W \in \mathbb{R}^{3 \times 4}$ ):

W = \begin{bmatrix} 5.0 & -2.0 & 1.0 & 4.0 \\ -5.0 & 5.0 & 5.0 & 0.0 \\ 5.0 & 0.0 & 0.0 & 4.0 \end{bmatrix} \begin{matrix} \text{(Row 1: Hit)} \\ \text{(Row 2: Oscar)} \\ \text{(Row 3: Franchise)} \end{matrix}

Matrix Dimension Formalism & Notation

In neural network mathematics, matrices are denoted using uppercase capital letters ( $W$ ), while vectors are denoted using lowercase letters ( $x, w, b$ ).

The shape of a matrix is defined as:

W \in \mathbb{R}^{m \times n}

Where:

$m$ is the number of rows (representing the number of parallel neurons, or outputs).
$n$ is the number of columns (representing the number of input features).

NOTE: Reading Dimensions: Rows First, Columns Second Matrix dimensions are always written as Rows $\times$ Columns (remember the order as RC). Here, $W \in \mathbb{R}^{3 \times 4}$ indicates 3 rows and 4 columns.

To reference a specific entry inside a matrix, we use two indices: $W_{ij}$ , where $i$ denotes the row index and $j$ denotes the column index:

W = \begin{bmatrix} W_{11} & W_{12} & W_{13} & W_{14} \\ W_{21} & W_{22} & W_{23} & W_{24} \\ W_{31} & W_{32} & W_{33} & W_{34} \end{bmatrix}

Row index $i$ selects the neuron (prediction): $i = 1$ is Hit, $i = 2$ is Oscar, and $i = 3$ is Franchise.
Column index $j$ selects the input feature: $j = 1$ is Action, $j = 2$ is Romance, $j = 3$ is Comedy, and $j = 4$ is Sci-Fi.

Reading individual weight entries:

$W_{11} = +5.0$ : The Hit neuron weight applied to Action.
$W_{21} = -5.0$ : The Oscar neuron weight applied to Action.
$W_{22} = +5.0$ : The Oscar neuron weight applied to Romance.
$W_{33} = 0.0$ : The Franchise neuron weight applied to Comedy.

By organizing weights into rows and columns, the matrix enforces our core axiom: Position is Identity (where position locks the semantic role of each number, distinct from an algebraic identity element). Every row pairs its columns with the exact locked input feature order [Action, Romance, Comedy, Sci-Fi], guaranteeing that each neuron performs consistent arithmetic across all incoming data.

To visualize how individual weight vectors stack into matrix rows to evaluate parallel predictions simultaneously, examine the matrix architecture below:

Weight Vectors Stacking into Parallel Neurons — Weight vectors stacking into rows ($W \in \mathbb{R}^{3 \times 4}$): Three distinct neurons (Hit, Oscar, Franchise) form horizontal rows that sweep across the 4-dimensional input feature vector $x$, evaluating multiple commercial outcomes in parallel.

Notice how the visual grid coordinates mirror our mathematical notation: each horizontal row corresponds to an individual neuron ( $i$ ), while each vertical column corresponds to an incoming feature dimension ( $j$ ). In Lesson 2, we will formalize the exact arithmetic that multiplies this 2D grid by the input vector $x$ to evaluate all three predictions in a single pass.