Assam HS Skill Subject Artificial Intelligence key terms

The field of building computer systems that perform tasks normally needing human intelligence, such as reasoning, learning, perceiving, and deciding. It is the broadest umbrella term covering machine learning, deep learning, NLP, and robotics.

A system that perceives its environment through sensors and acts on it through actuators to achieve goals. Agents may be reactive, goal-based, utility-based, or learning agents, and their behaviour is judged by rationality, not by human likeness.

A thought experiment proposed by Alan Turing in 1950: if a human interrogator cannot reliably tell a machine's text replies from a person's, the machine is said to exhibit intelligent behaviour. It tests conversational imitation, not genuine understanding.

A subfield of AI where programs improve at a task by learning patterns from data instead of following hand-written rules. A model is fitted on examples and then used to predict or classify new, unseen inputs.

Machine learning using neural networks with many stacked layers that automatically learn feature representations from raw data. It powers modern image recognition and language models but needs large datasets and substantial compute.

Training a model on labelled examples, where each input is paired with the correct output. The model learns a mapping from features to labels and is used for classification when outputs are categories and regression when outputs are numbers.

Finding structure in data that has no labels, for example grouping similar records or reducing dimensions. Clustering and association rules are typical tasks; there is no single correct answer to compare against during training.

Training an agent to choose actions by rewarding good outcomes and penalising poor ones through trial and error in an environment. Learning is driven by cumulative reward rather than by labelled examples.

A supervised task predicting a discrete category, such as spam or not-spam, from input features. Models are judged with accuracy, precision, recall, F1, and confusion matrices rather than error distances.

A supervised task predicting a continuous numeric value, such as price or temperature, from input features. Performance is usually measured with mean absolute error, mean squared error, or R-squared.

Training data is the portion of a dataset used to fit model parameters; test data is held back and used only to estimate performance on unseen inputs. Keeping them separate prevents an optimistic, misleading evaluation.

When a model memorises noise and quirks of the training set and therefore generalises poorly to new data. Symptoms include very high training accuracy with much lower validation accuracy; remedies include more data, regularisation, and simpler models.

When a model is too simple or poorly trained to capture the underlying pattern, so it performs badly on both training and test data. Adding relevant features, more capacity, or longer training usually helps.

A table of predicted versus actual classes showing true positives, true negatives, false positives, and false negatives. It reveals which classes a classifier confuses and underlies metrics such as precision, recall, and F1.

Precision is the share of predicted positives that are truly positive; recall is the share of actual positives that were found. They trade off against each other, and the F1 score is their harmonic mean.

Measuring how well a trained model performs, typically on held-out data, using metrics suited to the task. Good practice includes splitting data, cross-validation, and comparing against a simple baseline.

A resampling technique that splits data into several folds, training on some and validating on the rest in turn, then averaging the scores. It gives a more stable performance estimate than a single split.

A regression metric that averages the squared differences between predicted and actual values. Squaring penalises large errors heavily; taking its square root gives RMSE in the original units.

A feature is an input variable describing an example, such as age or pixel value; a label is the target answer attached to that example in supervised learning. Feature engineering is designing informative inputs.

An unsupervised technique that groups records so that items within a group are more similar to each other than to items in other groups. K-means is a common algorithm; the number of clusters often must be chosen.

A model built from layers of interconnected units, each computing a weighted sum of its inputs plus a bias and passing the result through an activation function. Stacking layers lets the network learn increasingly abstract features.

The simplest artificial neuron: it multiplies inputs by weights, adds a bias, and fires according to a threshold or step function. A single perceptron can separate linearly separable classes only.

A non-linear function applied to a neuron's weighted sum, such as ReLU, sigmoid, or tanh. Without non-linearity, stacked layers would collapse into a single linear transformation and could not learn complex patterns.

Weights scale how strongly each input affects a neuron; bias shifts the neuron's output independent of inputs. Both are learnable parameters adjusted during training to reduce loss.

The algorithm that computes how much each weight contributed to the loss by propagating errors backwards through the network, layer by layer. Those gradients are then used by an optimiser to update the weights.

An optimisation method that repeatedly nudges parameters in the direction that reduces loss, stepping by a learning rate. Too large a rate can diverge; too small a rate makes training slow.

An epoch is one full pass over the training dataset; a batch is the subset of examples processed before one weight update. Batch size affects memory use, training speed, and gradient stability.

The area of AI concerned with making text and speech machine-processable: tokenisation, cleaning, sentiment analysis, translation, and question answering. Modern NLP relies heavily on neural language models.

Splitting raw text into smaller units such as words, subwords, or characters so a model can process it. The choice of tokeniser strongly affects vocabulary size and how unknown words are handled.

Very common words such as 'the' or 'is' that are often removed before analysis because they carry little topical meaning. Removal reduces noise and size but can damage tasks where word order or phrasing matters.

A conversational program that interprets user messages and returns replies, either by matching rules or by generating responses with a language model. Design must cover intent handling, fallback replies, and safe boundaries.

The field that lets machines interpret images and video, for example by detecting objects or classifying scenes. Basic pipelines resize and normalise pixels, then feed them to a convolutional or other model.

Assigning one label, such as 'cat' or 'tumour', to a whole image. Performance is judged on held-out images, and accuracy can hide bias if some classes are far rarer than others.

A neural network that slides small learnable filters over an image to detect local patterns such as edges and textures, then pools and stacks these maps. Weight sharing makes CNNs efficient for grid-like data.

Detecting and fixing problems in a dataset: duplicates, missing values, inconsistent formats, outliers, and wrong types. Cleaning decisions must be documented because they shape every downstream result.

Filling gaps in data, for example with the mean, median, or a predicted value, or dropping affected rows. The choice can bias results, so the reason for each method should be recorded.

Summarising and visualising a dataset before modelling to reveal distributions, relationships, and anomalies. Techniques include histograms, scatter plots, correlation checks, and summary tables.

Numbers that summarise a dataset, such as mean, median, mode, range, variance, and standard deviation. They describe the observed data only and do not by themselves support causal claims.

An observation far from the rest of the data, caused by genuine extremes, measurement error, or data-entry mistakes. Outliers can distort means and model fits, so they should be investigated rather than deleted automatically.

Presenting data graphically so patterns are easy to see, using charts such as bar, line, scatter, and histogram. The chart type must match the data and the message, and axes should never mislead.

The ability to read, question, and communicate with data: understanding where numbers come from, what a chart really shows, and which conclusions the evidence supports. It includes healthy scepticism about sampling and correlation.

A distortion that occurs when the collected data does not represent the population it is meant to describe, for example surveying only smartphone users. Biased samples produce models that fail for under-represented groups.

Systematic unfairness in a model's outputs, often inherited from biased training data, unrepresentative features, or flawed objectives. It can deny loans or jobs to groups and must be tested for, not assumed absent.

Designing and deploying AI systems that are fair, transparent, accountable, safe, and privacy-respecting. It involves impact assessments, human oversight, clear documentation, and channels for affected people to challenge decisions.

Protecting personal information by collecting only what is needed, securing it, and using it only as the person agreed. Consent must be informed and revocable, and anonymisation must be genuine, not cosmetic.

Legal rules governing how personal data may be collected, stored, and shared, such as India's Digital Personal Data Protection Act, 2023. They typically require purpose limitation, security safeguards, and redress for data principals.

The degree to which a person can understand why a model produced a given output. Simple models are often inherently interpretable, while complex ones need post-hoc tools such as feature-importance explanations.

The shift of tasks from human workers to automation, which can remove some roles while creating others. Policy responses include retraining, social protection, and designing AI to augment rather than replace workers.

Turning a vague real-world need into a precise, testable AI task by defining the goal, users, inputs, outputs, and success metric. Poor framing is a leading cause of projects that are technically sound but useless.

The end-to-end sequence of framing a problem, collecting and preparing data, building and evaluating a model, deploying it, and monitoring it. Iteration is normal; teams often loop back to earlier stages.

Putting a trained model into real use and watching its behaviour over time. Monitoring detects data drift, performance decay, and harmful outputs, and triggers retraining or rollback when needed.

A list is an ordered, mutable sequence accessed by index; a dictionary stores key-value pairs accessed by key. Both are built-in Python types widely used to hold records and configuration before loading into pandas.

A fixed-type, multi-dimensional array from the NumPy library that supports fast element-wise arithmetic and broadcasting. It is the numerical foundation under pandas and most machine-learning libraries.

A two-dimensional labelled table from the pandas library with named columns and an index, supporting filtering, grouping, joining, and missing-value handling. It is the standard tool for tabular data preparation in Python.

Rescaling numeric features to a comparable range, for example to 0-1 or to zero mean and unit variance. It helps distance-based and gradient-based models converge and prevents large-magnitude features dominating.