AIP-210 Exam Prep
AIP-210 Exam Glossary - 40 Terms
Search the terminology pack for Certified Artificial Intelligence Practitioner. Use these definitions with the study guide and practice questions.
A
- Accuracy
- An evaluation metric measuring the proportion of total predictions that are correct.
B
- Backpropagation
- An algorithm for training neural networks that computes gradients of the loss with respect to model parameters by propagating errors backward through the network.
- Batch normalization
- A normalization technique that uses statistics computed across a mini-batch to stabilize and accelerate training.
- Bayesian optimization
- A sequential hyperparameter tuning method that uses a probabilistic model to choose promising parameter settings.
C
- Chain rule
- A calculus rule used in backpropagation to compute derivatives through compositions of functions.
- Class imbalance
- A data distribution problem in which some classes have far more examples than others.
- Convolutional filter
- A small learnable kernel applied across input data to extract features such as edges or textures.
- Convolutional Neural Network
- A neural network architecture designed for grid-like data such as images, using convolutions to detect local patterns.
D
- Decision tree
- A supervised learning model that makes predictions by recursively splitting data based on feature values.
- Dropout regularization
- A neural network regularization technique that randomly sets a fraction of activations to zero during training to reduce overfitting.
E
- Entropy
- A decision tree splitting criterion that measures uncertainty or disorder in class probabilities.
- Evaluation metric
- A quantitative measure used to assess model performance.
G
- GELU activation
- A smooth activation function commonly used in transformer models, applying probabilistic gating to inputs.
- Gini impurity
- A decision tree splitting criterion that measures node impurity based on class probability distribution.
- Gradient
- A derivative indicating how much a model parameter should change to reduce the loss function.
- Grid search
- A hyperparameter optimization method that evaluates all combinations from a predefined parameter grid.
H
- High-cardinality categorical feature
- A categorical variable with many unique values, often challenging for standard encoding methods.
- Hyperparameter tuning
- The process of searching for the best configuration settings of a machine learning model.
I
- Imbalanced binary classification
- A classification setting where one class heavily outnumbers the other, making some evaluation metrics misleading.
K
- K-fold cross-validation
- A resampling technique that splits data into k subsets and rotates the validation fold so each observation is used for both training and validation.
L
- Layer normalization
- A normalization method that normalizes across features within each individual sample, independent of batch size.
- Local connectivity
- A CNN property where neurons connect to small local regions of the input rather than the entire input.
- Low latency
- A system characteristic describing minimal delay between receiving input and producing output.
M
- MICE
- Multiple Imputation by Chained Equations, an iterative imputation method that models each feature as a function of the others.
- Model drift
- The degradation of model performance over time due to changes in data distributions or relationships.
N
- N-dimensional array
- A data structure for storing homogeneous numerical data across one or more dimensions efficiently.
- Neural network
- A machine learning model composed of interconnected layers of nodes that learn patterns from data.
- NumPy
- The foundational Python library for numerical computing, providing efficient n-dimensional arrays and vectorized operations.
O
- One-hot encoding
- A method of representing categorical values as binary indicator columns.
P
- Precision
- A classification metric defined as true positives divided by predicted positives, measuring correctness of positive predictions.
R
- Random search
- A hyperparameter optimization method that samples random combinations of parameters from specified ranges.
- Random Undersampling
- A class imbalance technique that reduces the majority class by randomly removing its examples.
- Reinforcement learning
- A learning paradigm in which an agent interacts with an environment and learns to maximize cumulative reward.
S
- scikit-learn
- A widely used Python library for classical machine learning on structured and tabular data.
- SMOTE
- A class imbalance method that creates synthetic minority-class examples by interpolating between existing minority samples.
- Streaming inference
- A deployment pattern in which incoming data is processed continuously with low latency for real-time predictions.
T
- Target encoding
- An encoding technique that replaces categorical values with statistics derived from the target variable, often with smoothing.
- Train/validation/test split
- A data partitioning strategy that separates data for model fitting, hyperparameter tuning, and final unbiased evaluation.
V
- Vectorized operations
- Array computations performed without explicit Python loops, typically optimized for speed.
W
- Weight sharing
- A mechanism in CNNs where the same filter parameters are reused across different input positions.
About These Definitions
These definitions are loaded from the shared release pack. Use them with the study guide and practice questions to connect vocabulary to exam scenarios.