AI · LESSON 09 / 25 · English
Neural Network Architecture
A neuron is a calculation, not a tiny human brain.
Content reviewed: 22 October 2026 · தமிழில் படிக்க ↗
Understand it in plain language
An artificial neuron combines inputs using weights, adds a bias, and applies an activation function. A network connects many such calculations. The word neuron is an analogy; the computation does not establish consciousness or human-style understanding. Learning adjusts parameters so a chosen loss improves on training examples.
How the AI approach works
A neural network is a parameterized function assembled from layers. A dense layer computes z = Wx + b. An activation such as ReLU introduces nonlinearity. Without nonlinearities between ordinary linear layers, stacking them still produces a linear mapping; depth alone would not provide the expressive behavior we usually want.
The input shape and output meaning are design decisions. A multiclass classifier can produce one logit per class, followed by softmax to obtain normalized scores. A binary classifier often uses one logit and a sigmoid. A regression head may output a real number without either transformation. Choose a compatible loss; a ReLU output is not automatically a probability.
Weights and biases are learned parameters. Width, number of layers, activation choices, and learning rate are design or training hyperparameters. Initialization matters because symmetric or poorly scaled weights can obstruct learning. Gradients passed through many layers can shrink or grow; normalization, residual connections, and appropriate initialization help address training difficulties in suitable architectures.
Read an architecture as a flow of tensor shapes, not a picture of a brain. Check which dimensions represent a batch, features, sequence positions, and classes. Count parameters to understand capacity and memory, then evaluate whether that capacity helps the actual task.
Break the idea into steps
- Multiply each input by its weight and add the results.
- Add a bias, which shifts the calculation even when inputs are zero.
- Apply an activation and trace how a changed input changes the output.
Worked example: connect the method to the task
For 4 input features, a hidden dense layer of 3 units has 4×3 weights and 3 biases: 15 parameters. A 2-class output layer adds 3×2 weights and 2 biases: 8 more. Total: 23 trainable parameters. The hidden activation enables a nonlinear boundary; the final logits describe the two supported classes.
Changing the number of classes changes the output head and labels, not just the displayed text. During training, backpropagation calculates gradients for these 23 parameters; inference simply uses their fitted values.
Apply the idea: check your reasoning
Think before revealing the answer
How many parameters does a dense layer with 5 inputs and 2 outputs have, including biases?
Reveal the answer and reasoning
5×2 weights plus 2 biases = 12. This count says nothing by itself about whether the model will generalize; that still requires an appropriate evaluation.
A common mistake: Not every network output is a probability. The task and output layer determine its interpretation.
Before you move on
- Explain the main idea without repeating the technical label.
- Identify the approach, its assumptions, and the evidence needed to trust the result.
- Complete the challenge, then compare your reasoning with the answer.
Sources & a deeper next step
The refresh uses primary papers and official documentation. These optional links are in English; you can understand this lesson without reading them first. Product versions and availability can change.