Begin with a simple relationship
Imagine a hypothetical model estimating exercises completed from study hours. We choose a weight of 2 and a bias of 1 only to illustrate the calculation; these numbers are not empirical findings or trained parameters.
- Input · x
- Input: study hours.
- Weight · w
- Weight: how much the estimate changes per extra hour.
- Bias · b
- Bias: a value added to the output.
For 3 hours: 2 × 3 + 1 = 7. This is the model’s prediction under our chosen parameters, not a promise about student performance.
Change the input. Observe the output.
Weight and bias stay fixed here. Changing an input is not training; training adjusts parameters using data and an error objective.
Where does learning begin?
With real examples of study hours and observed outcomes, we could compare predictions with observations and adjust parameters to reduce error. We must evaluate on data not used for training. An association does not establish causation, and our simplified example omits prior experience and exercise difficulty.
This is a linear-model example, not a complete neural network. Related operations appear inside neural units alongside other components such as activation functions.