The Basics of Artificial Neural Networks: What Urban Planners Should Know

This article is part of an ongoing column on AI and planning by urban planner and AI expert, Tom Sanchez. Read more installments here. 

Urban planners aren’t expected to become AI engineers. But with more planning tools powered by artificial intelligence (AI), and especially by a class of models called artificial neural networks (ANNs), it helps to know what’s happening under the hood. Even if we’re not the ones building these systems, understanding them gives us the ability to ask smart questions, spot potential issues, and make good decisions about how we use them. Understanding them is also important so that we can explain our methods and hopefully build trust in the process. 

The truth is, planners are already using tools that rely on ANNs, even if we don’t see them directly. Tools that drive chatbots, classify satellite imagery, forecast travel demand, and analyze public input all have ANNs working in the background. So, what are they? And how are they relevant to us?

What’s an ANN, anyway?

ANNs are a subset of methods within the sphere of AI with the broad goal of getting computers to perform tasks that typically require human intelligence. A major approach to achieving this is Machine Learning (ML), where computers learn from data rather than being explicitly programmed. Within ML, ANNs are models inspired by the brain that can recognize complex patterns. Deep Learning is a specialized form of ML that uses large, multi-layered ANNs to solve especially complex problems like image recognition or language translation.

At its core, an ANN is a type of mathematical model that learns patterns from data. It's inspired by neural networks in the human brain, but instead of neurons and synapses, these networks consist of layers of nodes and weighted connections. You can think of it like a multi-stage decision process. The first layer sees the raw data, which may be a row of numbers, pixels in an image, or words in a sentence. As the data moves through the network, each layer looks for patterns, usually based on many, many prior examples used to “train” the ANN (often also referred to as a “model”). 

A simple example

Without going into great detail on the mathematics involved, I’ll explain the basic elements of an ANN, including its nodes, layers, connections, and their functions.

The input layer is the starting point of an ANN. It’s where data enters the system. Each node (or "neuron") in the input layer represents a single feature or variable from the dataset (X₁, X₂, and X₃ in Figure 2, below). For instance, imagine you want to predict whether a parcel is residential or commercial based on just three variables: 1) parcel size (in square feet), 2) building height (in stories), and 3) percentage of vegetation cover. Your input layer would have three nodes (or variables), each holding one of those values. For a given parcel, it might look like:

X₁. Parcel size = 5,000

X₂. Building height = 2

X₃. Vegetation = 30%

These features are passed along to the next layer in the network via connections for further analysis. Think of the input layer as the front door of the model; it doesn’t change or process anything on its own; it just passes the raw data inside. All connections have calculated weights, shown as W₁, W₂, W₃, and W₄ in Figure 2. The ANN generated weights (i.e., levels of importance for each connection), are numerical values that the model learns and adjusts during training.

The hidden layers sit between the input and output layers, and they’re called “hidden” simply because they’re not directly visible, as they neither receive raw input nor produce final output, but instead, process and transform the data along the way. A simple ANN typically has one or two hidden layers. When a neural network has three or more hidden layers, it is generally referred to as a deep neural network, and the approach is known as deep learning, which is used for very complex patterns in images or speech.

Each hidden layer consists of nodes (or neurons), and each node is typically connected to every node in the previous and next layer. Each node applies a mathematical function to the weighted inputs it receives and passes the result forward. By matching outcomes from the training data and corresponding input values, the network calculates these weights and biases to fit the data (i.e., patterns). 

Suppose the network has one hidden layer with four nodes (as shown in Figure 2). Each of these four nodes takes a combination of the three input values, multiplies each by a learned weight, adds them up, and then applies an “activation function,” which is a mathematical function that determines thresholds of significance. Think of it like a valve that only allows positive signals or values to pass through, blocking the rest. The process of determining the weights, biases, and activation levels results from processing the data and matching to patterns in the training data. 

Hidden layers allow the network to transform raw data into meaningful intermediate patterns. For example, in our land use model, a hidden node might end up learning something like “is this parcel a large, low-rise building with lots of green space?” which is a combination of features that would likely indicate a residential use.

The output layer is the final step in a neural network. It takes the processed data from the last hidden layer and turns it into a result, such as a prediction, classification, or estimate, depending on the task and types of data. 

The output value is calculated by applying the set variables (X values), learned weights (W values), and biases (H values) to the results from the hidden layer, followed by a final activation function (the “valve” mentioned earlier). The output layer in this case would be a probability (value between 0 and 1) that a property is classified as residential (Y₁) and the probability that a property is commercial (Y₂). For the simple example above, a parcel that is 5,000 s.f., has a two-level structure, and 30% of the parcel is vegetated, may have Y₁ = 0.88 and Y₂ = 0.12. In other words, the ANN predicts there is an 88% likelihood that the given parcel is residential, and 12% likelihood that it is commercial.  

What this means

Even though complex math is buried inside the model, the key idea is that the output layer is where the model gives you its best prediction or estimation, based on all the patterns it’s detected in the input data (or training data). The more accurate the patterns, the more reliable the output — but only if the data is accurate, the task is suitable, and the model has been trained effectively. In a nutshell, using lots of example data, the neural network determines how best to combine and weight the input values to generate a “formula” that ultimately predicts outcome values or categories.

One challenge is that they are considered by some to be “black boxes.” While we can often see what the model predicted, it's not always clear how or why mathematically. But this is true with many optimization models that have many variables and many data points. The model can test a huge number of combinations to find the best possible “fit” to perceived patterns. 

Another issue is bias. Like other AI methods, neural networks learn from the data they’re fed. If that data reflects existing inequalities, such as in service coverage, law enforcement, or public investment, the network can learn and reinforce those patterns. If a predictive model trained on old policing data recommends more patrols in historically over-policed areas, for example, it could perpetuate systemic harm.

So, what should planners do?

You don’t have to understand the math behind neural networks, but it’s worth knowing the structure. Think of it like understanding how a traffic model works: even if you didn’t build it, you need to know its assumptions, inputs, and limitations.

When using AI tools powered by neural networks (or not), consider asking:

• What data was used to train this model?
• How accurate is it, and under what conditions?
• What patterns might it be missing or misrepresenting?
• Who might be excluded by this process?

The more we understand how these systems work, the better we can use and explain them, not just efficiently, but ethically and creatively. But as planners, in using these technologies, we bring something essential to the table: expertise and judgment. Our job is to ask the right questions, consider the human impact, and shape technologies to serve public goals. We don’t need to be technical experts. However, we must remain curious and informed.