Deep learning is a branch of machine learning that stacks many layers of artificial neurons to gradually turn raw input (pixels, sounds, words) into increasingly abstract concepts. The classic analogy: like a brain that first detects edges, then shapes, then a face, the network learns a hierarchy of representations, layer after layer.
The neuron, the basic building block
Each neuron computes a weighted sum of its inputs, adds a bias, then applies a nonlinear activation function. That nonlinearity (such as ReLU) is what lets the network model complex relationships.
$$ y = \sigma!\left(\sum_{i=1}^{n} w_i x_i + b\right) $$
The weights $w_i$ and bias $b$ are not hand-coded: they are learned automatically.
How the network learns
Training repeats two phases:
- Forward pass: data flows through the layers and produces a prediction.
- Backpropagation: the error is measured with a loss function, then gradient descent adjusts the weights to reduce that error.
Classical vs deep machine learning
| Criterion | Classical ML | Deep learning |
|---|---|---|
| Features | hand-crafted | learned automatically |
| Data required | moderate | massive |
| Compute | low | high (GPU) |
Deep learning now dominates vision, speech and natural language, and forms the foundation of today's large models.
Its strength: learning the right representations directly from data, rather than dictating them by hand.