How many hidden layers should a neural network have?

There is no single correct number of hidden layers for every neural network. The right depth depends on the complexity of the problem, the amount of data available, and how difficult the model is to train.

A simple rule of thumb

For relatively simple problems, one or two hidden layers may be enough.

For complex tasks such as image recognition, speech processing, or language modeling, networks may use many hidden layers.

Why use more hidden layers?

Additional hidden layers allow the network to build more complex hierarchical representations.

For example:

Pixels → Edges → Shapes → Object parts → Object

Each layer can transform the representation produced by the previous layer.

Why not keep adding layers?

More layers also increase:

• Computational cost
• Memory usage
• Training time
• Risk of overfitting
• Optimization difficulty

Very deep networks can also suffer from problems such as vanishing gradients or degradation if the architecture is not designed carefully.

How should the number of layers be chosen?

A practical approach is to:

• Start with a relatively simple architecture.
• Measure performance on validation data.
• Increase depth only when the model is underfitting or the task clearly benefits from deeper representations.
• Use techniques such as normalization and residual connections when training very deep networks.

The optimal depth is therefore usually found through experimentation rather than a fixed formula.

For the meaning of network depth itself, see “What does the depth of a neural network mean?”.

In short, a neural network should have enough hidden layers to model the task well, but not so many that unnecessary complexity makes training slower or less reliable.