MAI flashcards
MAI · Build recall with a focused review session.
32 cards. Read the term and the definition here. The deck at the top of this page is optional.
- Bias-variance tradeoff
- Expected prediction error decomposes into bias squared, variance, and irreducible noise. High bias underfits; high variance overfits. Regularization and more data reduce variance.
- Precision and recall
- Precision is true positives divided by predicted positives; recall is true positives divided by actual positives. F1 is their harmonic mean, balancing both.
- AUC-ROC
- Probability that a randomly chosen positive is scored above a randomly chosen negative. It is threshold-independent and equals 0.5 for random scoring.
- L1 versus L2 regularization
- L1 adds the sum of absolute weights and drives some weights exactly to zero, selecting features. L2 adds squared weights and shrinks all weights smoothly.
- No Free Lunch theorem
- No learning algorithm outperforms all others across every possible problem distribution. Algorithm choice must follow the data and the deployment loss.
- Stratified k-fold cross-validation
- Split data into k folds preserving class proportions, train on k minus one, validate on the rest, and average scores to estimate generalization.
- Data leakage
- Information from outside the training set, or from the target, contaminating model training. It inflates validation scores and causes production failure.
- Target encoding
- Replace a categorical value with the mean target for that category. Must be computed within training folds to avoid leakage.
- Missing data mechanisms
- MCAR means missingness is random; MAR means it depends on observed variables; MNAR means it depends on the unobserved value itself.
- Class imbalance evaluation
- Under imbalance, accuracy misleads because predicting the majority class scores high. Use precision-recall AUC, recall at fixed precision, or cost-weighted confusion matrices.
- ReLU activation
- Computes the maximum of zero and the input. It is cheap and reduces vanishing gradients but can produce dead units with zero gradient.
- Backpropagation
- Applies the chain rule through the computation graph to compute gradients of the loss with respect to every parameter, enabling gradient-based training.
- Batch normalization
- Normalizes layer activations per mini-batch to stabilize and speed training. Layer normalization instead normalizes per sample and suits transformers.
- Dropout
- Randomly zeroes a fraction of activations during training as regularization. It is disabled at inference, with scaling to preserve expected activation magnitude.
- Attention mechanism
- Weights inputs by softmax of scaled query-key similarity, then aggregates values. Multi-head attention learns several such weightings in parallel.
- Adam optimizer
- Adaptive optimizer tracking first and second gradient moments with bias correction, giving per-parameter learning rates. AdamW decouples weight decay.
- Autoregressive generation
- A model produces one token at a time, conditioning each new token on all previous ones. Decoder-only language models generate text this way.
- Temperature and top-p
- Temperature scales logits before softmax, controlling randomness. Top-p keeps the smallest token set whose cumulative probability reaches p.
- Retrieval-augmented generation
- Retrieve relevant documents by embedding similarity, insert them into the prompt, and generate a grounded answer. Reduces hallucination and enables citations.
- LoRA fine-tuning
- Parameter-efficient tuning that freezes base weights and trains small low-rank matrices added to layers, cutting memory and cost versus full fine-tuning.
- Hallucination
- Fluent model output unsupported by evidence. Mitigations include retrieval grounding, constrained decoding, self-consistency checks, and abstention on low confidence.
- Demographic parity
- A fairness criterion requiring equal positive prediction rates across groups. It can conflict with equalized odds and calibration, which cannot all hold together.
- SHAP values
- Shapley-value-based attributions assigning each feature a contribution to a prediction, with local explanations and additive global importance.
- Prompt injection
- An attack where untrusted input overrides model instructions, causing unsafe actions or data exposure. Defenses include input filtering and least-privilege tool access.
- Model card
- A governance document describing a model's intended use, training data, performance across subgroups, limitations, and ethical considerations for reviewers.
- EU AI Act risk tiers
- Regulatory framework classifying AI systems as unacceptable, high, limited, or minimal risk, with obligations scaling by tier and by provider or deployer role.
- Canary deployment
- Route a small traffic slice to the new model, compare metrics against the incumbent, then ramp up or roll back based on observed performance.
- Shadow deployment
- Run the new model on live traffic without serving its outputs, logging predictions to compare against the production model before promotion.
- Feature store
- Centralized system serving consistent features for training and inference, preventing train-serve skew and enabling reuse across models.
- Concept drift
- A change in the relationship between inputs and target, degrading a deployed model even when input distributions look stable. Requires retraining or adaptation.
- Population stability index
- A drift metric comparing binned feature distributions between reference and current data. Values above roughly 0.2 signal meaningful shift.
- AI project ROI
- Benefit minus total cost, divided by total cost. Total cost includes data work, compute, staffing, integration, and ongoing monitoring and retraining.