Artificial Intelligence Course Revision & Terminology

Vocabulary practice flashcards covering history, evaluation metrics, reinforcement learning, NLP, and computer vision from the Artificial Intelligence course.

Syllogism

A formal deductive reasoning structure introduced by Aristotle that derives valid logical conclusions from two or more propositions.

Turing Test

A test formulated by Alan Turing in 1950 to evaluate machine intelligence by testing whether a human evaluator can distinguish between text responses from a computer and a human.

Artificial Intelligence (John McCarthy)

The science and engineering of making intelligent machines, especially intelligent computer programs, a term coined by John McCarthy around the 1956 Dartmouth conference.

AI Winter

A period characterized by a major decline in research interest, public perception, and funding for artificial intelligence, primarily caused by unrealistic expectations and technical limitations.

Expert System

A knowledge-based system designed to emulate the decision-making and problem-solving process of a human domain expert using formalized rules and facts.

Knowledge Base

A primary component of an expert system that systematically stores formalized expert knowledge, rules, and facts from a specialized application domain.

Inference Engine

The core component of an expert system that processes the knowledge base using logical deduction rules to infer new facts, solutions, and recommendations.

Perceptron

An artificial neuron model invented by Frank Rosenblatt in 1957 that computes a weighted sum of inputs and applies an activation function to generate an output.

Backpropagation Algorithm

A mathematical method popularized by Rumelhart et al. in 1986 that allows artificial neural networks to learn by calculating error gradients and updating connection weights backward through layers.

Gartner Hype Cycle

A graphical representation of the life cycle and market maturity of emerging technologies, moving through Innovation Trigger, Peak of Inflated Expectations, Trough of Disillusionment, Slope of Enlightenment, and Plateau of Productivity.

Artificial Narrow Intelligence (ANI)

Also called Weak AI, systems designed to perform specialized problem-solving tasks within a single domain—often faster than humans—without the ability to transfer knowledge to other domains.

Artificial General Intelligence (AGI)

Also called Strong AI, a theoretical system possessing human-level cognitive capabilities across multiple domains, including abstract reasoning and true generalization capabilities.

Artificial Superintelligence

A hypothetical AI system that matches and dramatically exceeds human intellectual and cognitive abilities across all fields, often achieved via recursive self-improvement.

Robotic Process Automation (RPA)

The automated execution of repetitive, manual, rule-based, or time-consuming business workflow tasks by software bots.

Internet of Things (IoT)

A network infrastructure connecting physical and virtual devices to exchange data using information and communication technologies.

Development Set

Also known as a validation set, a dataset separate from the training set used to evaluate an AI model during development and optimize its hyperparameters.

Confusion Matrix

A performance table for classification models on test data with known true values, categorizing outcomes into True Positives, False Positives, True Negatives, and False Negatives.

Precision

The proportion of positive predictions that are truly correct, calculated as $\frac{TP}{TP + FP}$ .

Recall

The proportion of actual positive samples that were identified correctly, calculated as $\text{Recall} = \frac{TP}{TP + FN}$.

Accuracy (Classification)

The proportion of all correct predictions relative to the total number of evaluated samples, calculated as $\text{Accuracy} = \frac{TP + TN}{TP + TN + FP + FN}$.

F-Score

The harmonic mean of precision and recall, calculated as $F = 2 \cdot \frac{\text{precision} \cdot \text{recall}}{\text{precision} + \text{recall}}$.

Reinforcement Learning

A machine learning paradigm where an autonomous agent learns to maximize cumulative reward through iterative trial-and-error interaction with an environment.

Markov Property

The condition of a system where future state transitions depend solely on the present state and action, independent of preceding historical states.

Markov Decision Process (MDP)

A mathematical framework for modeling sequential decision-making composed of states $S$, actions $A$, transition probabilities $P_a(s, s')$, rewards $R(s, a, s')$, and a policy $\pi$.

Policy ($\pi$)

A decision-making strategy in reinforcement learning, formalized as $\pi(s, a) = p(a_t = a \mid s_t = s)$, specifying the action to take in a given state.

Value Function (RL)

A function that estimates the expected long-term cumulative discounted return of being in a given state under a specified policy.

Discount Rate ($\gamma$)

A factor $0 \le \gamma \le 1$ in reinforcement learning that weights the present value of future rewards, ensuring convergence over infinite horizons.

Temporal Difference (TD) Learning

A model-free reinforcement learning approach that updates value estimates based on the differences between consecutive predictions rather than waiting for terminal outcomes.

Q-Learning

A model-free temporal difference algorithm that iteratively optimizes the quality function $Q(s, a) = r + \gamma \max_{a'} Q(s', a')$ using the Bellman equation.

Exploration vs. Exploitation

The trade-off between trying random actions to discover new information about an environment (exploration) and leveraging known actions to maximize rewards (exploitation).

Tokenization

The natural language processing step that splits raw text into discrete smaller units, such as words, punctuation, or sub-words.

Lemmatization

An NLP preprocessing technique that uses grammatical vocabulary and morphological analysis to return words to their canonical dictionary base form (lemma).

Stemming

A rule-based text preprocessing technique that chops off suffixes and word endings to produce a base stem, sometimes yielding non-dictionary words.

Part-of-Speech (POS) Tagging

The process of marking words in a text with corresponding grammatical classes and syntactic categories such as nouns, verbs, adjectives, and pronouns.

Named Entity Recognition (NER)

The subfield of NLP focused on identifying and classifying specific entities in unstructured text into predefined categories such as person names, company names, locations, and dates.

Sentiment Analysis

The automated extraction and classification of subjective aspects from text, such as determining an author's emotional state, opinion, or sentiment polarities.

Pivot Machine Translation

A machine translation technique used for under-resourced language pairs where text is first translated into an intermediate bridge language before reaching the target language.

Bag-of-Words (BoW)

A text vectorization technique representing documents by word frequencies, ignoring word order, grammar, and semantic context.

Continuous Bag-of-Words (CBOW)

A Word2Vec neural architecture that predicts a missing target word using the surrounding context words within a sliding window.

Skip-Gram

A Word2Vec model variant that uses a single input word to predict the surrounding context words within a defined window.

TF-IDF

A statistical weighting measure in information retrieval that reflects the importance of a term $t$ in a document $d$ relative to a corpus $D$, calculated as $\text{TFIDF}(t, d) = \text{TF}(t, d) \times \text{IDF}(t)$.

GloVe (Global Vectors)

An unsupervised word vectorization algorithm combining global matrix factorization with local context window statistics using a word co-occurrence matrix.

Self-Attention Mechanism

A core transformer mechanism that calculates the relative significance and contextual relationships between all token positions in a sequence simultaneously.

BERT

Bidirectional Encoder Representations from Transformers; a pre-trained model developed by Google that learns contextual representations bidirectionally using Masked Language Modeling and Next Sentence Prediction.

Pixel

The smallest discrete element of a digital image array representing an individual intensity or color value.

2D Image Convolution

An image filtering method where a sliding matrix kernel is multiplied element-wise across local pixel neighborhoods to generate a modified output image.

Padding (Computer Vision)

Techniques such as constant, replication, or reflection padding used to add artificial pixel borders to input images to allow full kernel convolution operations.

Radial Distortion

An optical aberration causing straight lines in real-world scenes to curve near lens boundaries, encompassing barrel, pincushion, mustache, and fisheye types.

Tangential Distortion

An optical distortion occurring when a camera lens is not mounted perfectly parallel to the image sensor plane.

Intrinsic Camera Parameters

Internal optical parameters of a camera—such as focal length, optical center, and lens distortion coefficients—used to map 3D camera coordinates to 2D image coordinates.

Extrinsic Camera Parameters

Parameters describing the camera's external position and orientation (rotation and translation) in 3D real-world space.

Harris Corner Detection

A feature detection technique that shifts a small sliding window in all directions to find corners, identified where pixel intensity changes significantly in both horizontal and vertical directions.

BRIEF (Binary Robust Independent Elementary Features)

A computationally efficient local feature descriptor that encodes patch information into a binary string based on simple pixel intensity pairwise comparisons.

Semantic Segmentation

A computer vision task that assigns a class label to every individual pixel in an image to partition regions into semantic categories.

Variational Autoencoder (VAE)

A generative model composed of an encoder and decoder that maps inputs into a continuous, probabilistic latent space by optimizing the evidence lower bound (ELBO).

Generative Adversarial Network (GAN)

A generative framework introduced by Ian Goodfellow in 2014 where a generator neural network synthesizes data while a discriminator network evaluates whether data are real or synthetic.

Mode Collapse

A common challenge in training GANs where the generator produces only a very narrow variety of outputs, failing to capture the full distribution of the training data.

Diffusion Model

A class of generative models that synthesizes high-quality images through an iterative reverse denoising process, progressively refining pure noise into coherent images.

U-Net

A symmetric convolutional encoder-decoder architecture with skip connections between corresponding layers, primarily responsible for iterative noise removal in diffusion models.

CLIP (Contrastive Language-Image Pre-training)

A multimodal model developed by OpenAI that jointly trains image and text encoders on image-caption pairs to align visual and textual representations.