1/41
Comprehensive vocabulary flashcards covering Exam 2 concepts, including knowledge-based systems, neural network architectures, deep learning regularizations, model families, learning paradigms, generative and agentic AI, and model explainability techniques.
Knowledge-based system (KBS)
An AI program that stores human knowledge (facts and rules) in a knowledge base and uses reasoning to solve problems or give advice.
Expert system (ES)
A type of knowledge-based system that copies the decision-making of a human expert in a narrow area (e.g., MYCIN for diagnosing infections).
Knowledge-Based System Architecture
The structural design connecting the user and domain expert to components including the user interface, explanation facility, working memory, inference engine, knowledge base, and knowledge acquisition module.

Knowledge base
The component of a knowledge-based system that stores the facts and IF-THEN rules about the specific domain.
Inference engine
The 'brain' of a knowledge-based system that applies the rules to the facts to reach a conclusion using forward or backward chaining.
Working memory
The component of a knowledge-based system that stores the facts about the current case being evaluated.
Explanation facility
The component of a knowledge-based system that shows why and how the system reached its answer, including which rules were used.
Knowledge acquisition module
The component of a knowledge-based system through which the domain expert or knowledge engineer adds and updates knowledge.
Forward chaining
An inference method that starts from known facts and applies rules to move forward to a conclusion.
Backward chaining
An inference method that starts from a goal or hypothesis and checks backward to see if supporting facts exist.
Blackboard architecture
A shared workspace architecture where independent specialist modules (knowledge sources) post partial results and collaborate under a control component to solve a problem step by step.
Neural network (NN)
A machine-learning model loosely inspired by the brain, composed of connected nodes ('neurons') in layers that learn patterns from data by adjusting connection weights.
Neural Network Architecture
A layered structure consisting of an input layer (features), one or more hidden layers (weights + activation), and an output layer (prediction), interconnected by weighted links.

Backpropagation
A neural network training procedure where the error is sent backward through the network to calculate how much each weight contributed to the loss.
Deep learning
Machine learning using neural networks with many hidden layers that automatically extract increasingly abstract features directly from raw data.
Regularization
Techniques used in machine learning to reduce overfitting so the model generalizes well to new, unseen data.
Weight decay (L1 / L2 regularization)
A regularization method that penalizes large weights to keep the model simpler and prevent overfitting.
Dropout
A regularization technique that randomly switches off a subset of neurons during training so the network does not rely too heavily on any single unit.
Early stopping
A regularization method where training is halted as soon as the validation loss stops improving.
Data augmentation
A method to expand the training set by creating altered copies of existing data, such as flipping or rotating images.
Batch normalization
A neural network technique that stabilizes training and produces a mild regularizing effect.
Convolutional Neural Network (CNN)
A neural network that uses sliding filters across grid-like data to detect edges, shapes, and objects, commonly used for image recognition and medical imaging.
Recurrent Neural Network (RNN)
A neural network architecture with loops that pass memory from step to step, enabling it to process sequential data in order (e.g., LSTM).
Generative Adversarial Network (GAN)
A framework where two neural networks compete: a generator synthesizes fake data and a discriminator attempts to detect the fakes, improving both over time.
Transformer
A deep learning architecture introduced in 2017 that uses self-attention mechanisms to look at all tokens in a sequence concurrently and assess their relative importance.
Transfer learning
A technique where a model pre-trained on a large dataset or task is reused and fine-tuned for a new, related task, reducing data requirements.
Multitask learning
Training a single model simultaneously across multiple related tasks so that shared representations and insights benefit all tasks.
Reinforcement learning
A machine learning paradigm in which an autonomous agent learns behavior through trial and error by receiving rewards for positive actions and penalties for negative ones.
Supervised learning
Machine learning performed on labeled data (inputs with known outcomes) to perform classification or regression.
Unsupervised learning
Machine learning performed on unlabeled data (inputs only) to uncover hidden patterns, groupings, clustering, or dimensionality reduction.
Self-supervised learning
A training process where the data creates its own supervision signal, such as pre-training an LLM by hiding words within raw text and training the model to predict them.
Large Language Model (LLM)
A model trained on massive amounts of text with billions of parameters, whose primary function is to iteratively predict the most likely next word.
Generative AI
An umbrella term for AI systems designed to generate new content (such as text, images, video, audio, or code) by learning patterns from sample data.
Prompt engineering
The practice of crafting and structuring instructions for generative models to obtain high-quality and reliable outputs without altering model weights.
Retrieval-Augmented Generation (RAG)
An architecture that retrieves relevant excerpts from external reference documents and feeds them into the model's prompt to ground responses in factual evidence.
Agentic AI
An AI framework that combines a model with planning, tools, memory, and oversight to autonomously pursue multi-step goals and execute actions.
Explainability
The process of providing human-understandable reasons for a model's prediction; an activity performed around or after model inference.
Interpretability
An intrinsic property of a machine learning model describing how easily a human can comprehend its internal mechanics and logic (e.g., linear models, small decision trees).
Reproducibility
The characteristic of an experiment or model where using identical data, code, and configuration settings consistently produces the exact same results.
LIME (Local Interpretable Model-agnostic Explanations)
An explainability technique that slightly alters an individual input, observes how predictions change, and fits a simple local surrogate model to highlight important features.
SHAP (Shapley Additive exPlanations)
A game-theoretic explainability method that fairly calculates the marginal contribution of each input feature to an individual prediction or global view.
Counterfactual explanations
An explainability method that identifies the minimum necessary change to input features that would reverse or flip the model's output.