1.What a single artificial neuron does
The name "artificial neural network" is borrowed from the nerve cells of the brain, but what it actually does is a very simple calculation. A single artificial neuron takes several inputs, multiplies each by an "importance," adds them all up, and passes the sum through a filter before sending it on. This importance is called a weight, and the result of multiplying and adding is called the weighted sum.
The final "filtering" step is called activation. One of the most common versions outputs 0 if the sum is less than 0 and outputs the sum unchanged if it is 0 or more. This simple bend is what lets neurons stacked in several layers represent complex relationships that no single straight line can express.
- Weighted sum = 0.6 × 4 + 3 × 0.5 + (−2)
- = 2.4 + 1.5 − 2 = 1.9
- Activation: 1.9 is greater than 0, so 1.9 is passed on unchanged.
- If the chance of rain were 0.1 and you were out for 1 hour, you would get 0.4 + 0.5 − 2 = −1.1, and the output would be 0.
2.The weights are what it has learned
In the example above, the weight of 4 on the chance of rain means "the chance of rain strongly affects the decision," and the bias of −2 lowers the default toward "usually don't take an umbrella." If a person sets these numbers, it is a rule-based program; if the machine sets them by looking at data, it is learning.
Training a neural network ultimately means adjusting these weights and biases. As we saw in Lesson 2, you measure the gap between the prediction and the correct answer (the loss) and nudge every weight a little in the direction that reduces the loss. Real neural networks have anywhere from thousands of such weights to, in large models, well over hundreds of billions. That is why training takes so much data and computation.
The key point is that a trained neural network does not store sentences like "cats have pointy ears." What it has learned is spread across the values of countless weights. That is why it is hard for a person to explain why a neural network made the decision it did.
3.What changes when you stack layers
Placing several neurons side by side makes a layer, and passing one layer's outputs on as the next layer's inputs makes a multi-layer neural network. The layers between the input layer and the output layer are called hidden layers, and learning with a neural network that has many hidden layers is called deep learning. "Deep" means the layers go deep.
As the layers get deeper, the network can capture more and more complex features, step by step. When researchers analyze networks that recognize photos, they tend to find that early layers respond to simple patterns such as lines and edges, middle layers to parts such as eyes or wheels, and later layers to whole shapes such as faces or cars. No one told it to "find the edges"; the roles divided up that way during training.
| Position | What it mostly responds to | Analogy |
|---|---|---|
| Early layers | Lines, edges, boundaries between colors | Pen strokes of letters |
| Middle layers | Parts such as eyes, ears, wheels | Words |
| Later layers | Whole objects such as faces, cats, cars | The meaning of a sentence |
4.A tiny neural network, calculated by hand
To see what "stacking layers" actually means as a calculation, let's follow a tiny neural network by hand: 2 inputs → 2 hidden neurons → 1 output. The numbers are chosen for illustration. The hidden neurons use the activation from earlier (0 if negative, otherwise unchanged), and the output neuron passes its weighted sum on unchanged.
The order of calculation is always the same. The outputs of one layer become the inputs of the next, and each neuron just does "multiply, add, filter." Calculating in order from the inputs toward the output like this is called the forward pass. The second example is one step of training: looking at that result and adjusting a weight once.
- Step 1: Weighted sum for h₁: 1 × 0.5 + 2 × 1 + (−1) = 0.5 + 2 − 1 = 1.5. It is positive, so h₁ = 1.5.
- Step 2: Weighted sum for h₂: 1 × (−2) + 2 × 0.5 + 0 = −2 + 1 = −1. It is negative, so after activation h₂ = 0.
- Step 3: Output: 1.5 × 2 + 0 × 3 + (−1) = 3 + 0 − 1 = 2.
- Check: Because h₂ became 0, its weight of 3 had no effect on this result. With different inputs, h₂ could turn on and join in. Changing which neurons turn on depending on the input is exactly what activation does.
- Step 1: Current prediction: 3 × 2 = 6. The error is 6 − 10 = −4, and the loss is (−4)² = 16.
- Step 2: How much the loss changes when w changes slightly (the gradient) is 2 × error × input = 2 × (−4) × 2 = −16. It is negative, so increasing w reduces the loss.
- Step 3: New w = w − learning rate × gradient = 3 − 0.05 × (−16) = 3 + 0.8 = 3.8.
- Step 4: New prediction: 3.8 × 2 = 7.6. The error is −2.4, and the loss is (−2.4)² = 5.76.
- Check: The loss dropped from 16 to 5.76. Repeating the same method brings w closer to 5, which gives exactly the right answer (5 × 2 = 10).
5.A neural network is not a brain
Because of the name, it is easy to assume that neural networks copy the brain exactly. In reality they borrow only one idea from the brain, "many units connected to each other, exchanging signals," and work very differently. An artificial neuron is a formula of multiplication and addition, and its learning is a kind of mathematical optimization that differs from human experience.
So it is not accurate to say that a neural network understands or feels the way a person does. Then again, dismissing it as "just calculation, nothing special" is not accurate either. When simple calculations come together on an enormous scale, they produce results sophisticated enough to surprise people. Keeping both facts in mind together is a balanced understanding.
📌 Key points
- Artificial neuron = a calculation that adds up input × weight (the weighted sum) and filters it through an activation
- Training means nudging the weights and biases a little so that the loss goes down
- The calculation is a forward pass that repeats "multiply, add, filter" layer by layer from input to output; training looks at the loss of that result and adjusts the weights
- What it has learned is stored not as sentences but spread across the values of countless weights
- Learning with neural networks that stack many layers is deep learning, and each layer captures increasingly complex features
- A neural network is a mathematical model that borrows an idea from the brain, not a copy of the brain
🤖 Try asking AI like this
Copy a prompt and replace the [ ] parts with your own situation. Don't take the answer on trust — check it against this lesson.
When you want to learn the principle by calculating by hand
Using a tiny neural network with 3 inputs, 2 hidden neurons and 1 output neuron as an example, set the weights to simple whole numbers and show me step by step how one input gets calculated all the way to the output. Organize it in a table so I can follow along and do the calculation myself.
When you want to know how far the analogy holds
Make a table of three ways artificial neural networks are similar to the human brain and five ways they are different. Mark anything you're not sure about as "uncertain."
- Standard explanations in introductory deep learning textbooks (artificial neurons, activation functions, layers)
- High school "Fundamentals of Artificial Intelligence" course content (Korean national curriculum, Ministry of Education)
Reached every goal above? Mark the lesson complete.
Storage is unavailable in this browser, so this lasts only for this page.🤖 AI Basics
- 1What Is AI? — Rules and Learning
- 2Machine Learning Basics — Data, Model, Training, Evaluation
- 3Neural Networks, Intuitively — Small Calculations Add Up to Judgment
- 4How Do Large Language Models Produce Text?
- 5Prompt Basics — Saying Exactly What You Want
- 6Advanced Prompting — Roles, Examples, Steps, Format
- 7Hallucination and Verification — How to Doubt What Sounds Plausible
- 8Privacy, Copyright and Ethics — Using AI Responsibly
- 9Working with AI — Delegate, Check, Connect
- 10Learning Strategies for the AI Era — The Basics Shape Your Questions