AIF-C01: AI & ML Fundamentals

Accuracy Metric

Total Correct Predictions; Ratio of true results across all classifications

Precision Metric

Quality of Positive Predictions; True Positives / (True Positives + False Positives); Crucial when false positives are costly

Recall Metric

Quantity of Actual Positives Caught; True Positives / (True Positives + False Negatives); Crucial when missing positive cases is dangerous

F1 Score

Harmonic mean of Precision and Recall; Provides a balanced single metric for evaluating imbalanced datasets

Area Under the ROC Curve (AUC)

Imbalanced Class Separation; Distinguishing power across all classification thresholds

K-Nearest Neighbors (k-NN)

Supervised Classification Algorithm; Label assignments based on closest surrounding data points

K-Means Clustering

Unsupervised Clustering Algorithm; Groups unlabeled data structures into K clusters

Least-to-Most Prompting

Deconstruct Complex Tasks; Sequential problem breaking and sub-task ordering

Inference

Running new data through a trained model to generate predictions; Uses existing weights without retraining

Training

The process of adjusting model weights using a training dataset to learn patterns and minimize error

Supervised Learning

Training a model on labeled data (inputs paired with correct output targets; e.g. regression / classification)

Unsupervised Learning

Training a model on unlabeled data to discover hidden patterns; groupings; or structures

Overfitting

Model memorizes training noise; High accuracy on training data; poor performance on new unseen test data

Underfitting

Model is too simple to capture underlying data patterns; Performs poorly on both training and test data

Hyperparameters

External configuration settings set BEFORE training begins to govern the learning process (e.g. learning rate; batch size)

Low Bias and Low Variance

Ideal model performance; fits training data well and generalizes accurately to unseen test data; exam trigger: "well-balanced model" / "high accuracy on both train and test sets"

Low Bias / High Variance (High Variability)

Overfitting; model memorizes training data noise but fails on unseen test data; exam trigger: "performs well on training data / poorly on test data"

High (Increased) Bias / Low (Less) Variance

Underfitting; model is too simple to learn underlying patterns in either dataset; exam trigger: "performs poorly on both training and test data"

High Bias / High Variance (Increased Variability)

Worst-case model failure; model is both too simple to capture patterns and highly unstable across different datasets; exam trigger: "consistently poor and erratic performance across all data splits"

Association Rule Learning

Unsupervised Rule Mining; identifies strong relationships and co-occurrence patterns between items in large datasets (e.g. market basket analysis)

Logistic Regression

Supervised Classification Algorithm; predicts discrete categorical outcomes and probabilities using a sigmoid function

Continuous Regression

Supervised Numerical Prediction; predicts continuous numeric value outputs based on input feature relationships

Probability Density Estimation

Unsupervised Distribution Analysis; estimates the continuous probability distribution function of an underlying random variable dataset

Clustering

Unsupervised Grouping Technique; partitions unlabeled data points into clusters based on feature similarity and distance metrics

RAG (Retrieval-Augmented Generation)

Knowledge Retrieval Framework; connects foundation models to external verified data sources without full model retraining

BERTScore

Semantic Evaluation Metric; uses contextual embeddings to measure semantic similarity between model output and reference text

BLEU Score

Precision-Based NLP Metric; evaluates machine translation quality by counting exact n-gram matches against reference text

ROUGE Score

Recall-Based NLP Metric; evaluates text summarization quality by measuring overlapping n-grams and word sequences against reference summaries

Traditional ML Models

Structured Data & Manual Features; works with numerical / categorical rows and columns; requires manual feature engineering; uses standard algorithms (e.g. regression / decision trees)

Generative AI Models

Unstructured Data & Automatic Features; works with text / images / audio; automatically learns representations via neural networks; uses advanced architectures (e.g. GANs / Transformers) to generate new content

GenAI Capability: Adaptability

Task & Domain Flexibility; adapts seamlessly across various tasks and functional domains

GenAI Capability: Responsiveness

Real-Time Output; generates content and responses in real-time

GenAI Capability: Simplicity

Task Deconstruction; simplifies complex tasks and user workflows

GenAI Capability: Creativity and Exploration

Novel Content Generation; generates novel ideas; designs; or technical solutions

GenAI Capability: Data Efficiency

Small Sample Learning; learns patterns effectively from relatively small amounts of data

GenAI Capability: Personalization

Tailored User Experience; creates personalized content tailored to individual preferences or characteristics

GenAI Capability: Scalability

High-Volume Output; generates large amounts of content quickly at scale

ML Lifecycle: Business Goal Identification

Phase 1; defining business objectives; success criteria; and identifying key performance indicators (KPIs)

ML Lifecycle: ML Problem Framing

Phase 2; translating business objectives into specific machine learning tasks (e.g. classification; regression; clustering)

ML Lifecycle: Data Collection

Phase 3; gathering relevant historical datasets; data streams; or third-party data needed to train the model

ML Lifecycle: Data Integration and Preparation

Phase 4; combining; cleaning; deduplicating; handling missing values; and structuring raw data for model ingestion

ML Lifecycle: Feature Engineering

Phase 5; selecting; transforming; and constructing relevant input attributes from raw data to improve model predictive power

ML Lifecycle: Model Training

Phase 6; feeding processed data into machine learning algorithms to adjust weights / parameters and learn patterns

ML Lifecycle: Model Validation

Phase 7; evaluating model performance on unseen test data using statistical metrics (e.g. accuracy; precision; recall; RMSE)

ML Lifecycle: Business Evaluation

Phase 8; testing trained model outputs against business success criteria and ROI before deploying to production

ML Lifecycle: Production Deployment

Phase 9; deploying validated models to hosting endpoints (e.g. real-time inference / batch transform) for live predictions