A neural network is a computer model that learns patterns from examples. It is not a tiny brain, and it does not understand a task in the human sense. Instead, it repeatedly adjusts numerical values until its outputs become more useful for a particular objective.
Layers turn inputs into predictions
A network is made of connected units arranged in layers. The first layer receives numbers representing an input—for example, pixel values from an image. Middle layers transform those numbers, and the final layer produces an output such as a label or a probability. Each connection has a weight that influences how strongly one unit affects another.
Training is a loop of prediction and correction
During training, the model makes a prediction for an example and compares it with the expected answer using a loss function. The loss is a numerical measure of how far the prediction is from the target. An optimisation algorithm then adjusts the weights to reduce that loss. Backpropagation efficiently calculates how each weight contributed to the error, allowing the model to make a more informed adjustment.
The loop repeats across many examples. With suitable data and training, the network can learn patterns that help it make predictions for examples it has not seen before. That ability is called generalisation; simply memorising the training data is not enough.
Training and using a model are different stages
Training changes the model’s weights and can require substantial computing. Inference is what happens after training, when the model uses its fixed weights to produce an answer for new input. When an app classifies a picture or suggests text, it is typically running inference—not retraining the model on the spot.
What neural networks do not guarantee
A network’s output depends on the examples, objective and design used to train it. A confident prediction is not automatically a correct one, and patterns in the training data can carry through to the model. Good evaluation uses data kept separate from training and checks how the model performs on the cases that matter in practice.
In short, neural networks learn by tuning connections to reduce prediction error. That simple idea, scaled across many layers and examples, supports a wide range of modern AI—but careful data, evaluation and human judgement still matter.
Source: Google Machine Learning Crash Course
Questions or corrections? Write to contact@welkdock.com. Back to the Blog.
