Neural networks are the powerhouses driving modern machine learning, and they appear in almost every cutting-edge AI system.
For example, a neural network can be used to recognize handwritten digits. Such a system might learn up to 128 different “things” about these digits. This could seem like a random number, but there’s something really interesting going on there. Each of these things was learned by a neuron—a simple processing unit in our neural network. The term neuron was originally inspired by neurons in the human brain, but while biological neurons receive signals and decide whether to fire based on those inputs, our artificial version is more like a simplified mathematical model. We start with inspiration from biology, but we end up somewhere quite different. Instead of trying to replicate the incredibly complex machinery of the brain, we’ve developed these artificial neurons to follow mathematical principles that turn out to work remarkably well for learning patterns.
You’ll notice that we use the term neural networks rather than artificial neural networks because the modern machine learning community dropped the “artificial” qualifier a while ago. In the machine learning community, “neural network” refers specifically to computational models, not biological ones. Using artificial neural network or its abbreviation ANN is considered outdated terminology in professional machine learning contexts.
Let’s go back to our digit recognition example to see how this works in practice. Those 128 neurons were working together, each learning to recognize different aspects of the digits. One neuron might have learned to spot curves, another to detect straight lines, and others to recognize different angles or intersections. It’s like having 128 tiny experts, each focused on one specific feature. But the real magic happens when we combine all of these pieces of information. These 128 individual observations work together to make the final decision about which digit was shown in the image.
That’s the real power of neural networks—they can take many simple pieces of learned information and combine them into surprisingly complex and accurate decisions. And this is just the beginning of what they can do. In this section, we’ll explore how they work in more detail.
Let’s break down a neural network into its basic building blocks. Think of it like understanding how a car works—you don’t need to know every technical detail, but understanding the main components helps you grasp how everything works together. We’ve already met our first component: the neuron. Each neuron is a tiny decision-maker. Just as a person might decide whether to bring an umbrella by looking at the clouds, wind, and weather forecast, a neuron makes simple decisions based on the information it receives. In our digit recognition example, one neuron might look at a small part of the image and decide how likely it is to contain a curved line.
But neurons don’t work alone—they’re connected to each other in interesting ways. Imagine a group of art experts trying to identify a painting. One expert might focus on the color palette, another on the brushstrokes, and a third on the composition. They share their observations with each other before making a final decision. Similarly, neurons are connected to each other, passing information along these connections. When one neuron makes a decision, it shares that information with other neurons that it’s connected to. These neurons are organized into layers, like departments in a company. The input layer is like the reception desk—it receives the raw information. In our digit recognition system, this layer takes in the image of the handwritten digit. Next, come the hidden layers—think of them as different departments processing the information. Recall our 128 neurons. They formed a hidden layer, each learning to recognize different patterns in the digits. Finally, we have the output layer, like the executive team making the final decision. In our case, this layer decides which digit (0–9) it thinks it’s looking at. Let’s see how information flows through this whole system. Here’s the sequence when we show it a handwritten digit:
This flow of information, from input through hidden layers to output, is what allows neural networks to transform raw data (e.g., pixel values) into meaningful decisions (e.g., digit predictions). It’s similar to how a large organization processes information—from initial data collection through various departments’ analyses to final decision-making. See the figure below for how information flows from one layer to another in a neural network.
What makes this system powerful is how each component plays its part while working together with others. Individual neurons make simple decisions, connections ensure information flows where it’s needed, and layers organize this process into manageable steps. When we train the network, we’re really teaching all these components to work together effectively, like training a team to operate smoothly.
How does a neural network learn? The process is really fascinating and, in some ways, mirrors how we might learn a new skill. When a neural network first starts out, it’s like a blank slate with random guesses. Consider those 128 neurons we talked about, each one initially making arbitrary decisions about what patterns to look for in our handwritten digits. They’re like novice art critics who haven’t yet learned what to look for.
This is where the magic of supervised learning comes in. Remember how we talked about learning from examples where we know the correct answer? That’s exactly what happens here.
We show the network thousands of handwritten digits, each one labeled with the correct number it represents. At first, the network makes lots of mistakes. A messy 6 might be mistaken for an 8, or a slightly tilted 1 might be seen as a 7. But here’s the clever part: every time the network makes a mistake, it adjusts itself slightly. Those 128 neurons gradually refine what they’re looking for. One neuron might realize it should pay more attention to curves at the top of the digit. Another might learn to focus more on straight lines at the bottom. The connections between neurons also adjust learning which patterns are more important for identifying each digit. Every mistake is an opportunity to learn.
A parallel to this might be a human learning to identify birds. At first, you might just make random guesses. But after seeing many examples of different birds, along with their correct names, you start noticing important details. Gradually, your observations of the shape of the beak, the color patterns, and the way they fly become more refined and reliable. Neural networks learn in a similar way, but they can process thousands of examples much faster than we can. Speed is their superpower. The key is that learning happens incrementally. The network doesn’t suddenly jump from random guesses to perfect predictions. Instead, it gets a little bit better with each example it sees. Early on, the improvements might be dramatic—like finally realizing that all 1s tend to be tall and thin. Later improvements might be more subtle.
This is why having many good examples is so crucial. Just as we saw in supervised learning, the quality and quantity of training examples largely determine how well the system will learn. The more handwritten digits the network sees, and the more diverse these examples are, the better it becomes at handling new, unseen digits in the future. It’s a simple principle with powerful results.
Let’s clear up some common misconceptions about machine learning. These are ideas that have somehow taken root in the field but don’t tell the whole story.
First, there’s a persistent belief that neural networks are exclusively tools for supervised learning. You’ll hear this repeated in many introductory courses and even some textbooks. But it couldn’t be further from the truth. Neural networks are more like a Swiss Army knife—incredibly versatile tools that can be applied to all types of machine learning.
Another misconception is that unsupervised learning is just about grouping similar things together. While clustering is certainly a powerful application, it barely scratches the surface of what unsupervised learning can do. Neural networks have opened up entirely new possibilities in this space. They can learn to generate new images, understand the structure of language without any labels, and even discover patterns we never knew existed in our data.
These concepts might sound abstract right now, and that’s okay. You will get there if you put the time and effort in. Think of it like learning to cook. You might start by following recipes before you understand why certain ingredients work together or what chemical reactions are happening in your food. But as you keep cooking, these deeper insights naturally emerge from your hands-on experience.
Editor’s note: This post has been adapted from a section of the book Keras 3: The Comprehensive Guide to Deep Learning with the Keras API and Python by Mohammad Nauman. Dr. Nauman is a seasoned machine learning expert with more than 20 years of teaching experience and a track record of educating 40,000+ students globally through his paid and free online courses on platforms like Udemy and YouTube.
This post was originally published 10/2025.