Is every neural network deep?

No. Deep learning uses networks with multiple processing layers, while the broader family includes simpler forms.

Does a network need labeled data?

Not always. Supervised learning uses labels, while self-supervised and other approaches obtain learning signals differently.

Can a trained network explain its answer?

Some explanation methods can help inspect behavior, but a fluent explanation or visualization is not automatically a faithful account of the computation.

AI Glossary / Deep learning / Beginner

Neural network

Definition

A neural network is a computational model made from connected units that transform numerical information. Its adjustable parameters can be learned from data, allowing it to recognize patterns, estimate quantities, create representations or generate outputs.

What it means

The units are often called artificial neurons, and their connections carry numerical weights. The name reflects a loose inspiration from nervous systems; it does not mean the software is a biological brain.

Neural networks form a broad family. A simple network may have a small hidden layer, while a larger system may contain many specialized components. Convolutional networks, recurrent networks and Transformer-based models organize computation differently.

The network learns according to its training data and objective. Predicting a label, reconstructing an input and predicting the next item in a sequence teach different behaviors. Calling a system a neural network therefore explains only part of what it does. Its task, architecture and training process are also important.

How it works

Inputs must be represented numerically. These might be measurements, image values or representations of text. In a common unit, the model combines inputs using weights, adds a bias and applies an activation function. Multiple units and stages produce the final output.

During training, a loss function measures a chosen kind of error. In many networks, backpropagation computes gradients and an optimizer changes parameters to reduce that loss. Other neural-network families can use different learning rules.

During inference, the trained model processes new inputs. This usually does not change its learned weights unless an explicit updating procedure is running. Intermediate values or conversational context are not the same thing as retraining.

Performance is assessed on examples not used to fit or select the model. A network can fit its training data while failing on new conditions, so low training error alone is not enough.

Three practical examples

A quality-control model identifies visual patterns associated with known product defects in inspection photographs.
A forecasting model estimates demand from information available before the period being predicted.
A language model generates a draft by repeatedly producing outputs conditioned on the supplied context, with the draft reviewed for its intended use.

Think of it this way

Think of a large arrangement of adjustable dials that transforms incoming signals into a result. Training changes the dials using examples and a rule for measuring error. Running the trained arrangement uses the dial settings to process a new input.

The analogy separates learning from use, but it does not imply that each dial has a simple human-readable meaning. Many parameters work together, and a useful prediction can arise without the model holding an explicit explanation that a person would recognize.

Worked example

Seeing how one weight can change

Consider a deliberately simple output calculation y_hat = wx. Let x = 2, the target y = 5 and the current weight w = 2. The prediction is 4.

Use the teaching loss L = one-half × (y_hat − y)². The loss is 0.5, and its gradient with respect to w is (wx − y)x = (4 − 5) × 2 = −2.

With a learning rate of 0.1, a gradient-descent update gives w_new = 2 − 0.1 × (−2) = 2.2. The new prediction is 4.4, and the loss falls to 0.18.

This illustrates one local parameter update, not a complete multilayer network or proof of future accuracy. Larger networks contain many interacting parameters, and improvement on one example does not establish improvement on every other example.

Illustrative case study

A fictional delivery business, Maple Routes, trains a neural network to predict arrival delays. The team first uses every field in its historical records and obtains an impressive test score.

An operational review finds that one field records a status entered after the delivery ended. It would not exist when the prediction is needed. The team removes it, builds a time-appropriate evaluation and compares the model with a simpler estimate based on recent routes.

The revised result is less dramatic but more relevant. Dispatchers receive a prediction based on genuinely available information, and the team monitors changing conditions. The case shows why the surrounding data and workflow can matter more than the size of the network.

Common misconceptions and mistakes

A neural network learns automatically from every interaction. Updating learned parameters requires a defined training or adaptation process.
More parameters always mean better results. Data quality, objective, architecture and evaluation all affect usefulness.
Neural networks are the whole of AI. They are one important family among many approaches.
A correct-looking answer establishes understanding or truth. Outputs still require evaluation appropriate to the task.

Business use

Neural networks can be valuable when complex patterns in data support a useful prediction or representation. They are not a substitute for deciding what outcome matters and whether the available evidence supports it.

Business teams should specify the input information, target, evaluation population, cost of mistakes and action taken after an output. Compare with practical baselines and account for data preparation, inference expense, monitoring and retraining. A model that performs well in a benchmark may need a different workflow before it reliably supports an employee or customer. The model's role should remain clear: estimate, rank, generate or recommend.

Check your understanding

Answer: A deployed model processes 1,000 new requests, but no training procedure runs and its parameters are unchanged.

Has it necessarily learned from those requests?

Answer: No. Processing inputs is inference. Learning would require an updating mechanism; temporary intermediate values do not by themselves change the trained model.

Related terms

Learn next

Trace the worked parameter update with a different learning rate. Then distinguish the data used for training, validation and final testing in a simple project. Study one architecture suited to your information type rather than treating all neural networks as identical.

Sources

Google, Machine Learning Glossary, neural-network terminology: https://developers.google.com/machine-learning/glossary Goodfellow, Bengio and Courville, Deep Feedforward Networks: https://www.deeplearningbook.org/contents/mlp.html Goodfellow, Bengio and Courville, Representation Learning: https://www.deeplearningbook.org/contents/representation.html Scikit-learn, supervised neural-network models: https://scikit-learn.org/stable/modules/neural_networks_supervised.html

Previous: Neural modeling fields Next: Neural network Gaussian process