AIAI cheat sheet

AI is the broad field of systems performing tasks associated with human intelligence; ML is the subset that learns patterns from data rather than following hand-coded rules.

Supervised learning uses labeled data (input-output pairs); unsupervised learning finds structure in unlabeled data; reinforcement learning learns from reward signals via trial and error.

Narrow (weak) AI targets a specific task and is what exists today; general (strong) AI would match human performance across domains and is hypothetical.

Inference is applying a trained model to new inputs; training is the process of fitting model parameters to data.

Parameters are learned weights internal to a model; hyperparameters (learning rate, batch size, tree depth) are set before training and control the learning process.

Overfitting = low training error but high generalization error; underfitting = high error on both training and held-out data.

A model's prediction can be classified as true positive, true negative, false positive, or false negative relative to the ground-truth label.

Turing Test operationalizes intelligence as indistinguishability from a human interlocutor; it measures behavior, not internal understanding.

Standard lifecycle: problem framing → data collection → feature engineering → training → evaluation → deployment → monitoring → retraining.

Train/validation/test split: train fits parameters, validation tunes hyperparameters, test gives an unbiased final estimate and is used once.

Precision = TP/(TP+FP); Recall = TP/(TP+FN); F1 = harmonic mean of precision and recall.

Accuracy is misleading on imbalanced data; prefer precision, recall, F1, or AUC-ROC when class frequencies are skewed.

k-fold cross-validation rotates the validation fold k times and averages scores, reducing variance of the performance estimate on small datasets.

Regression metrics: MAE (mean absolute error), MSE/RMSE (penalizes large errors more), R² (proportion of variance explained).

Regularization (L1/Lasso, L2/Ridge) penalizes large weights to reduce overfitting; L1 can drive weights to zero and perform feature selection.

Model drift: data drift changes input distribution, concept drift changes the input-output relationship; both require monitoring and retraining triggers.

Data quality dimensions: accuracy, completeness, consistency, timeliness, validity, and uniqueness.

Missing data options: deletion (listwise/pairwise), mean/median/mode imputation, model-based imputation; deletion biases results if missingness is not random.

One-hot encoding creates a binary column per category; label/ordinal encoding assigns integers and implies order, so use it only for ordinal variables.

Normalization scales features to [0,1]; standardization rescales to mean 0 and standard deviation 1; distance-based models (k-NN, SVM, k-means) require scaling.

Data leakage = information from the target or test set contaminating training features; it inflates validation scores and breaks in production.

Outliers can be detected via z-scores (|z| > 3), IQR rule (outside Q1 − 1.5·IQR to Q3 + 1.5·IQR), or isolation forests.

Class imbalance remedies: resampling (over/under), class weights, and threshold tuning; resample only the training set, never validation or test.

Feature engineering creates derived variables (ratios, aggregations, date parts) that make relevant patterns easier for the model to learn.

Transformer self-attention lets each token weigh all other tokens in context, enabling parallel processing and long-range dependency capture.

LLMs are trained in stages: pretraining (next-token prediction on large corpora) followed by instruction tuning and preference alignment (e.g., RLHF).

Tokenization splits text into subword units; context window is the maximum number of tokens the model can attend to at once, input plus output.

Temperature controls sampling randomness (low = deterministic, high = diverse); top-p/top-k truncate the probability mass considered at each step.

RAG (retrieval-augmented generation) retrieves relevant documents and inserts them into the prompt, grounding answers without changing model weights.

Embeddings map text to dense vectors where semantic similarity corresponds to vector distance (cosine similarity, dot product).

Hallucination is fluent but factually unsupported output; it arises because models optimize token likelihood, not truth.

Fine-tuning updates model weights on task data; prompt engineering and RAG change inputs only and are cheaper but do not alter learned parameters.

Bias sources: historical bias in training data, sampling/selection bias, labeling bias, and proxy variables that correlate with protected attributes.

Fairness metrics conflict: demographic parity (equal positive rates), equal opportunity (equal TPR), and predictive parity (equal PPV) generally cannot all hold simultaneously.

Transparency vs. explainability: transparency discloses how a system works; explainability describes why a specific output occurred (e.g., SHAP, LIME).

Human-in-the-loop places a human decision-maker over consequential AI outputs; meaningful oversight requires the human to have authority and context to override.

Informed consent, purpose limitation, and data minimization are core privacy-ethics principles for personal data used in AI.

Model cards and datasheets document intended use, training data, performance across subgroups, and known limitations.

Accountability requires a named owner for each AI system; unclear ownership is a recurring governance failure.

Red-teaming probes a system adversarially for harmful outputs, bias, and safety failures before and after release.

Prompt injection: malicious text in input or retrieved content overrides system instructions; mitigations include input validation, privilege separation, and output filtering.

Data poisoning corrupts training data to implant backdoors or degrade performance; membership inference attempts to determine whether a record was in the training set.

Model extraction/stealing queries a deployed model to approximate its behavior; rate limiting and query monitoring reduce exposure.

Adversarial examples are inputs with small perturbations that cause misclassification; adversarial training improves robustness at some accuracy cost.

Privacy techniques: anonymization/pseudonymization, differential privacy (calibrated noise), and federated learning (model updates shared, raw data stays local).

Governance controls include data classification, access control, audit logging, model inventory, and documented approval gates.

Regulatory frameworks to know: EU GDPR (personal data), EU AI Act (risk-tiered obligations), and NIST AI Risk Management Framework (govern/map/measure/manage).

Encryption in transit (TLS) and at rest, plus secrets management, are baseline controls for AI pipelines and model endpoints.

Use-case selection rule: start with a clearly defined business problem and measurable success metric, then choose the technique - not the reverse.

Good early candidates are high-volume, repetitive, data-rich tasks with tolerance for error and a clear feedback signal.

Common patterns: classification (fraud/spam), regression (demand forecasting), ranking/recommendation, anomaly detection, and generative assistance (summarization, drafting).

ROI = (benefit − total cost)/total cost; total cost includes data, compute, integration, monitoring, and human review - not just model training.

Build vs. buy: buy/integrate when the capability is commodity and non-differentiating; build when proprietary data or workflow creates advantage.

Human-in-the-loop is required where errors carry legal, financial, or safety consequences (medical triage, credit denial, content moderation).

Process automation candidates: document processing, customer support triage, code assistance, and quality inspection from images.

Pilot-to-production failure is usually organizational (workflow integration, ownership, change management) rather than model accuracy.

MLOps = practices and tooling to version, test, deploy, and monitor models reproducibly across the lifecycle.

Serving patterns: batch (scheduled offline scoring), real-time online inference (low latency per request), and streaming (continuous event processing).

Model registry stores versioned artifacts with metadata and stage (staging/production) to enable rollback and audit.

Shadow deployment runs the new model alongside the old without serving its output; canary release sends a small traffic fraction to the new model first.

Monitoring must cover service health (latency, error rate), data quality (schema, drift), and model performance where ground truth is available.

CI/CD for ML adds data and model validation to the standard build-test-deploy pipeline; automated retraining triggers on drift or performance decay.

Containerization (e.g., Docker) plus orchestration (e.g., Kubernetes) packages the model with its dependencies for reproducible deployment.

Feature store centralizes feature definitions for training and serving, reducing train/serve skew.