START HERE

What you will learn

  • Follow a simple input-to-prediction calculation.
  • Explain weights, loss and training in ordinary language.
  • Understand why a prediction is not automatically an established fact.

Before you start: Basic arithmetic is enough for the example. You do not need to write code.

What is inside a neural network?

Inputs are numbers describing an example. They might represent image pixels, words converted into numbers or measurements such as the number of support tickets. A layer transforms those inputs. Many layers can be connected so that one layer’s outputs become the next layer’s inputs.

Weights are adjustable numbers that influence how inputs contribute to a calculation. A bias is an adjustable offset. Layers commonly include nonlinear functions, which allow combinations of layers to represent relationships more complex than a single straight-line calculation. “Neural” is a name for this model family; you do not need to imagine a human mind inside it.

What does learning mean here?

Training repeatedly compares model outputs with a training objective and changes parameters such as weights. A loss function turns the difference into a number the training procedure can try to reduce. Gradients describe how a small change in a parameter would affect the loss; an optimiser uses that information to choose an update.

The objective matters. A model trained to predict a category is not automatically optimising the business decision you care about. Better performance on its training examples does not establish that it will work on new examples. You need a separate evaluation.

Step-by-step process

  1. Represent an example numerically

    Choose inputs relevant to the task and prepare them consistently. In a ticket-effort example, inputs could include ticket count and a defined complexity measure.

  2. Calculate an output

    Pass the inputs through the model. This is the forward pass: calculations move from inputs towards a prediction.

  3. Compare with the training target

    If the example has a known result, calculate the loss. The choice of loss determines which differences training is trying to reduce.

  4. Update the parameters during training

    Calculate gradients and let the optimiser adjust parameters. Repeating this across examples is how the model is fitted. This step is not normally performed simply because you ask an already-trained model a question.

  5. Check new examples

    Measure performance on examples not used to fit the model. Look at the kinds of errors, not just the average result.

FOLLOW ONE EXAMPLE

Worked scenario

A deliberately simplified prediction of work hours

This arithmetic example uses one linear calculation to illustrate weights. It is not a realistic effort estimator or a complete deep neural network. Suppose the inputs are 10 tickets and complexity 2.

  1. Use illustrative weights of 0.2 for ticket count and 0.5 for complexity, with a bias of 1. The calculation is (10 × 0.2) + (2 × 0.5) + 1 = 4 predicted hours.
  2. Suppose the recorded outcome was 6 hours. With squared error as the loss, the penalty is (4 − 6)² = 4. That is a loss value, not “4% inaccurate”.
  3. Training would adjust parameters across many examples. You would then assess the resulting model on separate tickets rather than judging it from this one calculation.

What this shows: You can see where a prediction comes from, how a target creates a training signal and why evaluation is a separate job.

The process at a glance

During training, the prediction is compared with a target and parameters are updated. Using a trained model to make a prediction does not require that update loop.
ProcessDecisionHuman checkpoint
Read Mermaid diagram source
flowchart LR
 A[Numerical inputs] --> B[Layer calculations]
 B --> C[Prediction]
 C --> D[Compare with training target]
 D --> E[Calculate loss and gradients]
 E --> F[Update weights during training]
 F --> B

What you should have at the end

  • A correct explanation of inputs, weights and outputs.
  • An understanding of the difference between calculating a prediction and fitting a model.
  • Questions to ask about training objectives, evaluation data and relevant errors.

Common mistakes and how to avoid them

Treating weights as stored human-readable facts

Weights influence calculations; they are not a tidy table of verified statements.

Calling low training loss proof of usefulness

The model must also be evaluated on appropriate new examples.

Assuming every task needs custom training

An existing model, a simpler calculation or a better process may be sufficient.

APPLY THE IDEA

Try it yourself

Your task: Repeat the illustrative calculation using 20 tickets and complexity 2, leaving the weights and bias unchanged.

What to produce: The predicted hours and a plain-language explanation of which part of the calculation changed.

Check your understanding: You should obtain 6 hours. Can you explain why this arithmetic result does not prove the estimator is accurate?

Key terms in plain English

Parameter
An adjustable value inside a model, such as a weight or bias.
Layer
A stage of calculations that transforms one representation into another.
Loss
A numerical penalty expressing the training objective.
Gradient
Information about how a parameter change would affect the loss.

Sources and further reading