AI vs Machine Learning vs Deep Learning

Understanding AI vs Machine Learning vs Deep Learning

Different approaches suit different kinds of problems. A rule based program follows instructions written by people. It can be very reliable when the situation is clear, such as checking whether a password has enough characters or calculating a bus fare.

It struggles when rules become too numerous or hard to describe. Recognising a cat in a photo is difficult to write as a list of rules because lighting, angle, fur colour, and background can all change. A learning system handles this by finding useful patterns in many examples.

This does not mean it understands a cat in the human sense. It has learned statistical clues that often appear with the correct label.

Training usually starts with a dataset. Each example contains inputs and, in supervised learning, a known answer called a label. For a house price model, inputs might include floor area, location, and age.

The label is the sale price. The model makes an initial prediction, compares it with the label, then changes its internal settings to reduce the error. This cycle happens many times.

One way to measure the error is mean squared error. It finds the difference between each prediction and the real answer, squares each difference, then averages the results.

Squaring makes large mistakes count more heavily. After training, the important test is performance on new data that was not used for learning.

Deep neural networks learn through connected layers of simple calculations. Early layers can detect basic features, while later layers combine them into more detailed patterns. In image work, one layer may respond to edges, a later layer may respond to shapes, and later parts of the network may help identify an object.

During training, the model sends information forward to make a prediction. It then sends error information backward through the layers to adjust connection strengths. This process is called backpropagation.

More layers can represent complicated relationships, but they create many settings to learn. That is why deep models often need large datasets, powerful hardware, and careful training.

Students meet these ideas in recommendation feeds, spam filters, voice typing, translation tools, camera face detection, and medical image research. The result can seem certain even when it is only a prediction based on past examples. Data quality matters as much as model choice.

Biased, incomplete, or incorrectly labelled data can produce unfair or inaccurate results. A model can even memorise its training examples instead of learning a general pattern. This is called overfitting.

Watch for the difference between training accuracy and test accuracy. A strong model performs well on unfamiliar cases, not only on the examples it has already seen.