Does every neural network need labeled examples?

No. Supervised learning uses labels or targets, but neural networks also appear in self-supervised, unsupervised, and reinforcement-learning setups.

Is a neural network the same as deep learning?

No. Deep learning is a subset of neural-network methods involving multiple representation layers.

Does training error measure real-world usefulness?

Only partially. It describes fit to the training objective and data; separate evaluation is needed for new cases and operational requirements.

AI Glossary / Deep learning / Beginner

Artificial neural network

Definition

An artificial neural network is a computational model made from connected units that transform numerical information using adjustable parameters. Training changes those parameters so the network performs a task, such as estimating a value or recognizing a pattern. The name reflects a loose inspiration from biological nervous systems; an artificial neural network is not a complete simulation of a brain and does not inherently understand its task as a person would.

What it means

A neural network learns a mapping between a representation of its inputs and its outputs. An image may become pixel values, a text passage may become numerical vectors, and a business record may become selected measurements. What the network can learn depends partly on what those representations preserve.

The network's architecture determines how its calculations connect. A simple feedforward network passes information through successive layers. Other architectures use different connections and operations. “Artificial neural network” is therefore a broad category. Deep learning uses networks with multiple layers of learned representation, but not every neural network is deep.

How it works

In a basic unit, inputs are multiplied by weights, added with a bias, and passed through an activation function. Combining many such operations allows the network to represent more complicated relationships. Nonlinear activations are important because stacking only linear transformations remains equivalent to a linear transformation.

For supervised training, examples include targets. The network produces predictions, a loss measures their discrepancy from the targets, and an optimizer adjusts parameters. Backpropagation commonly supplies gradients for those adjustments.

After training, inference applies the learned computation to new inputs. Validation and test data help assess performance beyond the examples used for fitting. Training and inference are related but distinct stages.

Three practical examples

1. Image sorting: a model trained on labeled product photos suggests categories for new photos, with evaluation covering lighting, viewpoints, and products absent from the training set.

2. Demand estimation: a network maps a defined set of historical and contextual features to a numerical forecast. The team compares it with a simpler baseline and evaluates future periods appropriately.

3. Message routing: numerical representations of support requests feed a classifier that proposes a destination team. Staff review ambiguous cases and track whether changes in customer vocabulary affect routing quality.

Think of it this way

Think of a chain of adjustable mixing desks. Each desk combines incoming signals and transforms them before passing them onward. Training adjusts the controls so the final output better matches the task. The analogy helps explain composition and tuning, but the controls do not represent conscious reasoning, and there is no guarantee that each internal signal corresponds to a concept people can name.

Worked example

Consider a tiny network with two inputs, x1 = 2 and x2 = 1. Its first hidden unit computes 0.5 × x1 + 1 × x2 − 1 = 1. Its second computes −1 × x1 + 2 × x2 + 0.5 = 0.5.

Apply ReLU, which replaces negative values with zero. Both hidden values remain unchanged. An output unit then computes 3 × 1 − 2 × 0.5 + 1 = 3.

If the target for this example is 4, the squared error is (4 − 3)² = 1. Training would use this error, together with the computation's derivatives and the selected optimizer, to change parameters. The calculation is one forward pass and loss evaluation. It does not establish that these hand-chosen weights have been trained or that the network works well on other inputs.

Illustrative case study

Fictional case: Shelf Signal, a small catalog team, prototypes a neural classifier for product images. Its initial test accidentally includes photographs of the same items used during training, creating an overly reassuring view of performance.

The team rebuilds the evaluation around genuinely separate items and checks photographs taken under different conditions. It also adds a review path for uncertain or unsuitable inputs. No accuracy or commercial result is claimed. The example shows why a working training program is only one part of a useful system: the data split, operational environment, and handling of failures determine what the demonstration actually proves.

Common misconceptions and mistakes

A neural network is not automatically the best solution because a task is labeled “AI.” Simpler models may be easier to validate, cheaper to operate, or equally effective.

More layers and parameters do not guarantee better generalization. A network can fit training data while learning fragile or irrelevant relationships.

A high output score is not automatically a calibrated probability. Its interpretation depends on the model, training objective, and evaluation. Likewise, changing inputs at prediction time does not normally retrain the weights. Finally, biological vocabulary such as “neuron” should not be mistaken for evidence that the system possesses human awareness or judgment.

Business use

Business value comes from a measured improvement in a specific workflow. Before training, define the prediction target, the information available at decision time, and the costs of different errors. These choices determine whether an apparently strong model answers the right question.

Evaluate the full system, including input preparation, response time, maintenance, and human review. Keep a record of the data and model version used for each release. If the environment changes, previously measured performance may no longer describe current behavior. A model should therefore enter operations with monitoring and a clear response to degraded results, not merely a favorable training chart.

Check your understanding

A network predicts the training examples accurately but fails on new product photographs. Has it demonstrated a useful general solution?

Answer: No. The result shows that training fit alone is insufficient. Investigate generalization, data leakage, representation, and differences between the training and operational inputs.

Related terms

Learn next

Recalculate the tiny network with one changed input, then study the distinction between training, validation, and testing. Build a simple baseline before adding network complexity and inspect individual errors, not only an average score.

Sources

Primary implementation-learning sources: Google, “Neural networks: Nodes and hidden layers,” https://developers.google.com/machine-learning/crash-course/neural-networks/nodes-hidden-layers ; TensorFlow, “Basic classification,” https://www.tensorflow.org/tutorials/keras/classification . Both were opened to check architecture and training/evaluation context. The toy network, business examples, and fictional case are original teaching material; they are not claims of measured performance from those tutorials.

Previous: Artificial narrow intelligence Next: Artificial precision