Definition
Deep learning is machine learning based on neural networks with many successive layers. Training adjusts the network's parameters so later layers can build useful representations from the outputs of earlier ones. In an image model, for example, early layers may respond to local edges while later representations support object-level decisions, though the model is not required to follow that exact progression.
"Deep" is relative, not a certification with one agreed layer count. The useful distinction is that the system learns a stack of representations through optimization rather than relying only on features selected by people. Depth can improve expressiveness, but architecture, data, objective, optimization, and compute determine whether that capacity becomes useful.
What training changes
Deep networks are usually trained with gradient-based optimization and backpropagation. The same model family may be trained through supervised labels, self-supervised prediction, reinforcement signals, or combinations of objectives. The training method and dataset matter as much as the number of layers when interpreting a capability claim.
Distinguish it from nearby terms
- Machine learning also includes linear models, decision trees, kernel methods, and other approaches that are not deep neural networks.
- A neural network can be shallow. Deep learning names the multilayer regime, not every network.
- Representation learning is the broader goal of learning useful features from data. Deep networks are one way to do it.
- A foundation model is defined by broad training and downstream adaptability. Most current foundation models use deep learning, but the terms describe different properties.
Operational significance
Deep models can reduce manual feature engineering while increasing dependence on training data, optimization choices, hardware, and empirical evaluation. Their internal representations are learned rather than written as explicit rules, so tests, monitoring, and failure analysis must cover the operating distribution instead of assuming that depth supplies understanding.
Check your understanding
A boosted decision-tree model may outperform a neural network on a structured business dataset. Both are machine learning. Only the layered neural network is deep learning, and the label alone does not make it the better model.