Unit ID: DLPY-M01-U01 Estimated active time: 25-35 minutes

Start from familiar machine learning

In classical machine learning, you often prepare features yourself.

For example, in a tabular problem you may create:

• age groups;
• counts;
• ratios;
• date features;
• missing-value flags; and
• category encodings.

Then a model such as logistic regression, a tree, a forest, or boosting learns patterns from those prepared features.

Deep learning changes the balance. The model can learn more of the representation itself.

What representation means

A representation is a useful way to express the input.

For an image, raw pixels are not very helpful by themselves. A neural network may learn early patterns such as edges and colour changes, then later patterns such as shapes or object parts.

For text, raw characters or tokens are not enough. A model may learn vector representations that place related words, phrases, or documents closer together in a learned space.

For tabular data, deep learning can learn interactions, but it does not automatically beat classical methods. Many tabular problems still work very well with simpler models.

Layers and parameters

A neural network is built from layers. Each layer transforms the data.

The trainable numbers inside the model are parameters. During training, the model changes those parameters to reduce loss.

Simple version:

1. The model makes a prediction.
2. The loss function measures how wrong it is.
3. Backpropagation computes gradients.
4. The optimizer updates parameters.
5. The process repeats.

You do not need heavy calculus to begin. You do need to know what each step is responsible for.

What deep learning is good at

Deep learning can be useful when:

• inputs are high-dimensional, such as images, audio, or text;
• the useful representation is hard to hand-design;
• enough data or transfer learning is available;
• errors can be measured carefully;
• compute cost is acceptable; and
• the task has a clear evaluation method.

What deep learning is not good at by default

Deep learning is not automatically better when:

• data is small;
• labels are unreliable;
• the task needs clear simple explanation;
• the consequence is high;
• the problem can be solved by rules or a small model;
• the team cannot evaluate errors; or
• the runtime cost is too high.

Classroom rule

Use this rule:

> Deep learning is worth trying when representation learning is likely to help and the evidence process is strong enough to catch failure.

Both parts matter. If representation learning is useful but evidence is weak, the result can still be unsafe. If evidence is strong but deep learning adds no value, use a simpler method.

Mini practice

For each task, write try later, try now, or do not use deep learning.

1. Classify small product photos into five categories using 6,000 labelled images.
2. Predict loan approval using a small spreadsheet and unclear labels.
3. Detect whether a short support message is about billing, login, or course access, with a labelled safe dataset and a simple baseline.
4. Copy values from one reporting table to another.

There is no prize for choosing deep learning. The prize is choosing responsibly.

Mark complete