08. Common AI Terminology

AI has a dense vocabulary. The same concept sometimes has multiple names. This glossary gives you the precise definition for each term and how they relate.

---

Term	Definition
Model	A mathematical function mapping inputs to outputs, learned from data
Parameters / Weights	Numerical values inside a model adjusted during training
Training	The process of adjusting weights to minimize prediction error
Inference	Using a trained model to make predictions on new data
Dataset	Collection of examples used to train, validate, or test a model
Feature	An input variable used by the model (e.g., age, price, word)
Label / Target	The correct output the model should predict
Loss / Cost	A number measuring how wrong the model’s prediction is
Gradient Descent	Optimization algorithm that updates weights to reduce loss
Backpropagation	Algorithm to compute how each weight contributes to the loss
Epoch	One complete pass through the training dataset
Batch Size	Number of examples processed in one training step
Learning Rate	How large each weight update step is

---

Term	Definition
Overfitting	Model memorizes training data, poor on new data
Underfitting	Model too simple, poor on both training and new data
Generalization	Model performs well on unseen data
Bias (statistical)	Systematic error — model consistently wrong in one direction
Variance	Model too sensitive to training data — inconsistent across samples
Regularization	Techniques to prevent overfitting (L1, L2, dropout)
Cross-validation	Evaluating model on multiple train/test splits to reduce variance
Baseline	Simple model to beat — sets minimum acceptable performance

---

Term	Definition
Neural Network	Layers of interconnected nodes loosely inspired by the brain
Neuron / Node	Single unit that computes a weighted sum + activation
Layer	Group of neurons (input, hidden, output layers)
Activation Function	Non-linear function applied per neuron (ReLU, sigmoid, tanh)
Deep Learning	Neural networks with many hidden layers
CNN	Convolutional Neural Network — specialized for images
RNN	Recurrent Neural Network — specialized for sequences
Transformer	Architecture using self-attention for parallel sequence processing
Attention	Mechanism that weighs the importance of different input parts
Embedding	Dense vector representation of discrete items (words, images)

---

Term	Definition
Token	Unit of text the model processes (word, sub-word, or character)
Context Window	Maximum tokens a model can process at once
Prompt	Input text given to an LLM
Completion / Response	Model’s output
Temperature	Controls randomness of output (0 = deterministic, 1+ = creative)
Top-p / Top-k	Sampling strategies controlling output diversity
Hallucination	Model confidently generates false information
Fine-tuning	Further training a pretrained model on domain-specific data
RLHF	Reinforcement Learning from Human Feedback — used to align LLMs
RAG	Retrieval-Augmented Generation — grounding LLM with external docs

---

Term	Definition
Supervised Learning	Training with labeled examples (input → correct output)
Unsupervised Learning	Finding patterns without labels
Reinforcement Learning	Learning via reward/penalty signals
Train Set	Data used to fit the model
Validation Set	Data used to tune hyperparameters
Test Set	Data held out for final evaluation only
Data Drift	Distribution of input data changes over time
Class Imbalance	One output class much rarer than others (e.g., fraud)

---

Q: What is the difference between a parameter and a hyperparameter?

A: Parameters (weights) are learned from training data — they are what the model optimizes. Hyperparameters are settings you choose before training — like learning rate, batch size, number of layers, or dropout rate. Parameters are internal to the model; hyperparameters are external knobs you tune.

---

Q: What is the difference between overfitting and underfitting?

A: Overfitting happens when a model learns training data too well — including noise — and fails to generalize. It has low training error but high test error. Underfitting happens when a model is too simple to capture the real patterns — high error on both training and test data. The goal is the sweet spot between them (good generalization).