WBCHSE HS Artificial Intelligence & Data Science key terms

A system that perceives its surroundings through sensors and acts on them through actuators in order to reach a goal. Examples range from a thermostat to a self-driving car.

A thought experiment proposed by Alan Turing in 1950: a human judge chats with an unseen machine and a person, and must decide which is which. Passing suggests human-like conversation, not genuine understanding.

The classic approach in which a programmer writes explicit if-then rules that transform inputs into outputs. It suits well-defined tasks such as payroll, but struggles where rules are too complex to state by hand.

NumPy's core n-dimensional array: a block of same-type values stored contiguously, with shape and dtype attributes. Whole-array arithmetic on ndarrays runs in compiled code, making it far faster than Python lists.

A two-dimensional pandas table with labelled rows and named columns of possibly different types. Methods such as head(), groupby() and fillna() make it the standard container for tabular data analysis in Python.

Rewriting code so that operations apply to whole arrays at once instead of element-by-element Python loops. NumPy and pandas do the looping internally in fast compiled code, giving shorter and much quicker programs.

Filling gaps in a dataset instead of dropping rows, for example with the column mean, median or most frequent value, or by predicting from other columns. The choice affects results, so record why it was made.

An observation lying far from the rest of its distribution, often flagged using the IQR rule or z-scores. Outliers may be genuine rare events or data-entry errors, so inspect before deleting.

A chart that groups numeric values into bins and shows the count or proportion in each bin as adjacent bars. It reveals the shape of a distribution: centre, spread, skew and possible multiple peaks.

A chart plotting one variable on each axis with a dot per observation, used to see relationships, clusters and outliers between two numeric variables. A rising cloud suggests positive association.

Summary values describing a typical observation. The mean averages all values and is pulled by extreme scores; the median is the middle value and resists them; the mode is the most frequent value.

A measure of how far values typically sit from the mean, equal to the square root of the variance. A small value means tightly clustered data; a large value means wide spread.

A symmetric bell-shaped distribution described by its mean and standard deviation, in which about 68% of values fall within one standard deviation of the mean and about 95% within two.

A number between -1 and +1 summarising the strength and direction of a straight-line relationship between two variables. Zero means no linear association; it does not prove that one variable causes the other.

Systematic error caused when the collected sample is not representative of the population, for example surveying only people with internet access. Results then fail to generalise, no matter how large the sample.

Training a model on labelled examples, where each input comes with the correct answer, so it can predict the label for new inputs. Regression and classification are the two main forms.

Finding structure in data that has no labels, such as grouping similar customers into clusters or reducing many dimensions to a few. There is no correct answer to compare against, so results need interpretation.

A supervised task predicting a continuous number, such as price, temperature or demand, by fitting a relationship between input features and the target. Error is usually measured with MAE, RMSE or R-squared.

A supervised task assigning each input to one of a fixed set of categories, such as spam or not spam. Models output a class or a probability, and are judged with accuracy, precision, recall and confusion matrices.

When a model learns the training data too closely, including its noise, and so performs well on training examples but poorly on new ones. Remedies include more data, simpler models, regularisation and cross-validation.

Dividing a dataset into one part used to fit the model and a separate part used to estimate how it will perform on unseen data. The test part must never influence training.

A table of predicted versus actual classes showing true positives, true negatives, false positives and false negatives. Precision, recall and F1 score are calculated from its four counts.

The simplest artificial neuron: it multiplies each input by a weight, adds a bias, and passes the sum through a step function to give a binary output. It learns by adjusting weights after each mistake.

A non-linear function applied to a neuron's weighted sum, such as ReLU, sigmoid or tanh, deciding what it passes on. Without non-linearity, stacked layers would collapse into a single linear transformation.

The algorithm that trains neural networks: compute the loss, then use the chain rule to send error gradients backwards through the layers and update each weight in the direction that reduces loss.

A layer of neurons between the input and output layers of a network. Its units build intermediate features, and adding layers deepens the network so it can model more complex patterns.

Splitting raw text into smaller units called tokens, usually words or sub-words, and often lowercasing and stripping punctuation first. Every later NLP step, from bag-of-words to embeddings, works on tokens.

Representing a document by the counts of each vocabulary word, ignoring word order and grammar. It is simple and fast but loses context and produces sparse, high-dimensional vectors.

A weighting scheme where a word's score rises with its frequency in a document and falls with the number of documents containing it, so common words count for little and distinctive words stand out.

The NLP task of deciding whether a piece of text expresses positive, negative or neutral opinion, often by classifying reviews, tweets or feedback with a model trained on labelled examples.

The smallest picture element in a digital image, holding one colour value. A greyscale image stores one number per pixel; a colour image stores red, green and blue intensities, usually from 0 to 255.

A sliding filter operation that multiplies small patches of an image by learned weights and sums them, producing a feature map. Stacked convolutions detect edges first, then textures, then whole objects.

Assigning a label, such as cat, tumour or defective part, to a whole image. Modern systems learn features with convolutional networks trained on thousands of labelled pictures.

Creating extra training images by randomly rotating, flipping, cropping or adjusting brightness of existing ones. It reduces overfitting and helps a model cope with variation it will meet in real use.

Systematic unfairness in a model's outputs, often because its training data under-represents or mislabels some groups. It can deny loans or jobs to protected groups, so data and outcomes must be audited.

Removing or transforming personal identifiers such as names, addresses and ID numbers so individuals cannot reasonably be identified. It limits harm from data leaks and is often required before sharing datasets.

Telling people clearly what data will be collected, why, and who will see it, then obtaining their voluntary agreement. Consent must be specific and can be withdrawn, and it is central to lawful data use.

Using AI systems to perform tasks previously done by people, which can cut costs and create new roles while displacing some workers. Debate centres on reskilling, job quality and widening inequality.

Inspecting a dataset before modelling: checking shape, types and missing values, then summarising distributions and relationships with statistics and plots to guide cleaning and feature choices.

The full workflow of a project: define the question, collect and clean data, explore it, engineer features, train and evaluate models, then deploy and monitor results, iterating as needed.