Letter D

Information that computers can store, examine or learn from.

Data can be words, images, sounds, numbers, clicks or sensor readings. Its quality, legality and representativeness strongly affect what an AI system can do.

#

Creating varied training examples by making sensible changes to existing data.

It is like practising the same skill under slightly different conditions. A picture might be cropped, brightened or flipped so the model learns the important subject rather than one exact photograph.

#

The principles used to decide whether collecting, sharing and using data is responsible and fair.

Something can be technically possible and legally permitted while still raising ethical concerns. Data ethics asks whether people have meaningful choice, whether the use is proportionate, who benefits, who may be harmed and whether communities have a voice in decisions about their information.

#

Searching large collections of data for useful patterns or relationships.

The mining comparison suggests digging through a lot of material to find something valuable. It can reveal trends and connections, but a pattern is not automatically an explanation or proof of cause.

#

Using data, statistics and computing to understand problems and support decisions.

Data science can include collecting and cleaning information, finding patterns, making charts, running experiments and building predictive models. Not every data-science project uses AI.

#

Providing clear information about where data came from and how it was collected, changed and used.

It is the data equivalent of an ingredient label and supply history. Useful transparency describes sources, permissions, missing groups, cleaning steps and limitations in language that affected people can understand, rather than merely stating that 'data was used.'

#

Building and improving systems by treating data as a central engineering concern.

This approach pays close attention to how data is collected, labelled, checked, maintained and used throughout a system's life. Better data practices can improve an AI system more than repeatedly changing the model.

#

An organised collection of data used for training, testing or analysis.

A dataset might be a spreadsheet, a library of images or a huge collection of text. Datasets can contain gaps, errors, copyrighted material or private information, so their origins matter.

#

A prediction method that follows a branching series of questions.

It works like a flowchart: take one branch if the answer is yes and another if it is no, then continue until reaching a result. Trees are often easier to inspect than large neural networks.

#

Machine learning that uses neural networks with many layers.

Deep learning is behind many advances in language, images and speech. The layers learn increasingly useful patterns from large amounts of data instead of relying only on hand-written rules.

#

Convincing fake media made or altered with AI.

Deepfakes can imitate a person's face, voice or actions. They have creative uses, but can also enable fraud, harassment and misinformation, especially when presented without clear labelling.

#

A generative model that learns to turn random noise into a useful image, sound or other output.

During training it learns how to reverse a gradual noising process. When generating, it starts with noise and repeatedly refines it toward something matching the prompt.

#

A digital representation of a real object, place or process that changes as the real one changes.

A digital twin uses measurements and models to help people monitor, test or predict what may happen in its physical counterpart. It is more than a static 3D picture because it is connected to real or regularly updated data.

#

The part of a GAN that learns to tell real training examples from generated ones.

A discriminator acts like a forgery detective. The generator presents increasingly convincing fakes, and the discriminator tries to expose them. Feedback from the detective helps the forger improve, while better fakes force the detective to become sharper.

#

Training a smaller model to imitate useful behaviour from a larger one.

The smaller 'student' aims to keep much of the larger 'teacher' model's ability while becoming cheaper or faster to run. Some quality is often traded for efficiency.

#

A surprising pattern where a model gets worse as it grows, then starts getting better again.

The traditional expectation is that an overly complex model will overfit and keep performing badly. In some modern systems, test performance follows two downward slopes: it improves, worsens around a troublesome middle size, then improves again when the model becomes much larger. It is an observed pattern, not a promise that making any model bigger will help.

#