ASSEB Class 10 AI & Robotics key terms

The field of building computer systems that perform tasks normally needing human intelligence, such as recognising images, understanding language, planning, and making decisions. Systems may use rules, learned models, or both.

An AI system designed for one specific task or a small set of related tasks, such as face unlock or spam filtering. It cannot transfer its skill to unrelated problems the way a person can.

The labelled or unlabelled examples used to fit a model's parameters. Its size, balance, and quality largely determine how well the resulting model behaves on new, unseen inputs.

Data held back from training and used only to estimate how well a finished model performs on unseen examples. It must never influence model choices, or the estimate becomes too optimistic.

A learning approach where each training example comes with the correct answer (a label), and the model learns a mapping from inputs to those labels, for example classifying emails as spam or not spam.

Learning from data without correct answers provided. The model looks for structure such as groups of similar items or unusual outliers, for example grouping customers by buying habits.

A supervised task where the model assigns an input to one of a fixed set of categories, such as 'cat', 'dog', or 'bird'. Output is a discrete label, often with a confidence score.

A supervised task where the model predicts a continuous numeric value, such as tomorrow's temperature or a house price, rather than a category. Accuracy is judged by how close predictions are.

When a model memorises details and noise of its training data and performs poorly on new data. Signs include very high training accuracy but much lower validation accuracy.

A model's ability to perform well on inputs it has never seen before, not just on its training examples. It is the real goal of machine learning and is measured with held-out data.

A table comparing a classifier's predicted labels with true labels, showing counts of true positives, true negatives, false positives, and false negatives. It reveals which classes the model confuses.

Precision is the share of positive predictions that are correct; recall is the share of actual positives the model found. They trade off, so both are reported together.

An unsupervised technique that groups data points so items within a group are more similar to each other than to items in other groups, for example segmenting shoppers by behaviour without predefined categories.

A measurable input variable used by a model, such as pixel brightness, temperature, or word count. Choosing and scaling good features strongly affects model performance.

Rescaling numeric features to a common range, often 0 to 1, so that no variable dominates learning simply because its raw values are larger. It also helps training converge faster.

Systematic error in data collection or labelling that makes a dataset unrepresentative, leading a model to perform unfairly or inaccurately for some groups. It can enter through sampling, history, or labelling choices.

The handling of personal information so that collection, storage, and use respect consent, minimise exposure, and follow applicable rules. In AI projects it affects what data may be gathered and retained.

Presenting data as charts or graphs so patterns, trends, and outliers become easy to see. The chart type should match the question, for example a line chart for change over time.

A data point far from the rest of the observations. It may indicate a measurement error, a rare event, or a fault in a sensor, and should be investigated before being removed.

The branch of AI concerned with extracting information from images and video, including detecting objects, recognising faces, reading text, and estimating depth or motion.

Assigning a whole image to one category, such as 'ripe' or 'unripe' fruit. It differs from object detection, which also reports where each object is located.

A neural network that slides small learnable filters across an image to detect local patterns such as edges and textures, then combines them into higher-level features. It is widely used in vision tasks.

The area of AI that enables computers to process and generate human language, covering tasks such as translation, sentiment analysis, summarisation, and chatbots.

Splitting text into smaller units, such as words, subwords, or characters, before analysis. The choice of units affects how a model handles punctuation, rare words, and languages without spaces.

Automatically classifying the attitude expressed in text, for example positive, negative, or neutral, in product reviews or social posts. Sarcasm and context often cause errors.

A programmable machine that senses its environment, processes information, and acts physically, often through actuators, to carry out tasks with little or no human intervention.

A device that converts a physical quantity such as light, distance, temperature, or rotation into an electrical signal a controller can read. Sensors give a robot information about itself and its surroundings.

A component that converts an electrical command into physical movement or force, such as a motor, servo, or solenoid. Actuators are how a robot affects the world.

A motor packaged with feedback control so its shaft moves to and holds a commanded angle. Commonly used for robot joints and steering because position is set directly rather than by timing.

Rapidly switching a digital output on and off so the average voltage can be varied. It is used to control motor speed, LED brightness, and servo position without an analogue output pin.

A sensor that emits a sound pulse and measures the time until its echo returns, giving distance to an obstacle. Readings can be unreliable on soft, angled, or very thin surfaces.

A sensor module combining accelerometers and gyroscopes, often with a magnetometer, to report acceleration, rotation rate, and orientation. Used for balance, heading, and motion tracking; readings drift over time.

Combining readings from several sensors to produce a more accurate and reliable estimate than any single sensor gives, for example merging an IMU with wheel encoders to track a robot's position.

Adjusting a sensor or its readings against a known reference so outputs are accurate and consistent. It removes systematic offset and scale errors before data is used for control or training.

A single chip containing a processor, memory, and input/output pins that runs one program repeatedly to read sensors and drive actuators. Common in embedded robot controllers.

A dedicated computer built into a device to perform a specific control function, usually with tight limits on memory, power, and timing, such as the controller inside a robot vacuum.

The number of independent ways a robot or mechanism can move, such as joints that each rotate or slide. More degrees of freedom give more flexibility but make control harder.

A behaviour in which a robot detects nearby objects and steers or stops to prevent collision. It typically combines distance sensors with a control rule that reacts faster as obstacles get closer.

The aim of designing and auditing AI systems so decisions do not systematically disadvantage groups of people. It involves checking outcomes across groups and questioning what the target label really measures.

The principle that identifiable people or organisations remain responsible for an AI system's design, deployment, and harms. It requires clear ownership, documentation, and routes for redress rather than blaming the model.

A design where a person reviews, approves, or can override an automated decision, especially in high-stakes settings such as medical screening or robot operation near people.

A hardware or software safeguard that prevents dangerous operation unless conditions are met, such as cutting motor power when a guard opens or an emergency stop is pressed.

Using an already trained model to produce predictions on new input. It is usually far cheaper than training and is what runs on a robot or phone in the field.

A requirement that a system produce its output within a fixed deadline, such as a robot updating motor commands many times per second. Missing the deadline counts as a failure, even if the answer is correct.