Module 3 · Lesson 6

Artificial neural networks were historically inspired by ideas about biological neurons, but modern neural networks are mathematical systems, not digital copies of the brain.

At their core, they repeatedly perform a simple kind of operation: combine numbers using learned weights, add a bias, apply a nonlinear transformation, and pass the result onward.

A single artificial neuron

inputs x
   ↓
weighted combination: w · x + b
   ↓
activation function
   ↓
output

Weights control how strongly different inputs influence the result. A bias gives the unit additional flexibility. An activation function introduces nonlinearity so that many layers together can represent relationships far more complex than a single straight-line transformation.

Layers build representations

A basic network is often described with an input layer, one or more hidden layers, and an output layer.

Input → hidden layer → hidden layer → output

In image recognition, early layers may become sensitive to local visual patterns while later layers combine those signals into more useful representations. In language models, the internal representations are much more complex, but the general principle remains: successive transformations build increasingly useful features for the objective the model is trained on.

Where is the model's “knowledge”?

It is common to say a model stores knowledge in its weights. That is a useful shorthand, but it can be misleading if taken too literally. A specific fact is usually not sitting in one clearly identifiable weight.

What the model has learned is distributed across large numbers of parameters and their interactions. The same parameter may contribute to many behaviors.

Why deep learning became so important

Deep networks can learn useful representations directly from raw or lightly processed data, reducing the need to hand-design every feature. Combined with large datasets, powerful hardware, improved optimization, and architectures such as transformers, this has produced major advances in language, vision, speech, and multimodal systems.

The key idea is not that neural networks imitate the human brain. It is that large compositions of learned mathematical transformations can represent remarkably complex patterns.

Demystifying AI series

← Reinforcement Learning   ·   What Is a Corpus? →