Machine Learning Tutorial

1. 1Machine learning starts with a baseline and a valid prediction clock
2. 2Machine learning core concepts: features, labels, validation and decisions
3. 3Vector vs raster in machine learning: geometry, pixels and feature vectors

Work through one boundary at a time and verify the example against real data.

1. 1Prediction-time feature availability: reject future information before training
2. 2Group and time validation: split by the failure you expect in production
3. 3Leakage-safe preprocessing: fit every learned transform inside the training fold
4. 4Decision thresholds: choose an action from probabilities and error costs
5. 5Probability calibration: test whether risk scores mean what they say

Compare simple baselines, inspect regression tails, regularize carefully and evaluate rare events and tree leaves.

1. 1Regression baselines and honest holdout metrics
2. 2Regression error slices and costly tails
3. 3Ridge regularization with training-only scaling
4. 4Rare-event precision, recall and changing prevalence
5. 5Decision-tree splits and minimum leaf support

Keep the final test untouched, inspect feature reliance and evaluate distance-based segments under a frozen transform.

1. 1Hyperparameter search and an untouched final test
2. 2Permutation importance on held-out data
3. 3Distance scaling before clustering
4. 4K-means objective and restart stability
5. 5Principal components and variance retention

Bagging, forest splits, boosting, learning curves, paired uncertainty and delayed-label monitoring.

1. 1Bootstrap bagging and out-of-bag evaluation
2. 2Random forest feature subsampling and leaf support
3. 3Gradient boosting residuals and early stopping
4. 4Learning curves and training-size diagnosis
5. 5Paired bootstrap intervals for model gain
6. 6Feature drift and delayed-label monitoring

Logistic loss, margin classifiers, density clusters and hierarchical linkage with separate evaluation contracts.

1. 1Logistic regression, log loss and odds
2. 2Linear margin classifiers and hinge loss
3. 3DBSCAN core, border and noise points
4. 4Agglomerative linkage and cluster cuts

Local neighbors, smoothed Bayes scores, fold-local feature choice, weighted loss and multiclass dispatch cost.

1. 1Nearest-neighbor distance and local support
2. 2Naive Bayes smoothing and dependent features
3. 3Fold-local feature selection and stability
4. 4Class weighting versus decision thresholds
5. 5Multiclass confusion and action costs

Self-training, weak rules, active review, delayed online updates and conditional prior adjustment.

1. 1Pseudo-label selection and contamination control
2. 2Weak-label rules, coverage and conflicts
3. 3Active learning with uncertainty and diversity
4. 4Online updates with delayed feedback
5. 5Label-prior shift and odds adjustment

Separate interval calibration, operating-slice coverage, group errors and review capacity.

1. 1Split conformal intervals for clearance forecasts
2. 2Prediction interval coverage by operating slice
3. 3Group error gaps and policy audit
4. 4Selective prediction and review capacity
5. 5Ensemble disagreement and outcome noise

Diagnose support, estimate risk under an explicit assumption, replay time-correct changes and compare a shadow candidate.

1. 1Covariate shift and support overlap
2. 2Importance-weighted risk under covariate shift
3. 3Concept change with mature outcomes
4. 4Rolling-origin retraining with an outcome embargo
5. 5Champion–challenger shadow comparison

Log decisions, evaluate alternative actions offline and review exploration and reward maturity.

1. 1Contextual bandit action logging and support
2. 2Inverse-propensity policy value
3. 3Augmented propensity policy value
4. 4Exploration budget and action guardrails
5. 5Delayed bandit rewards and outcome maturity

Pin the source and target contracts, test frozen representations, tune carefully and audit target regressions.

1. 1Pretrained encoder and target-task contract
2. 2Frozen embeddings and a linear probe
3. 3Staged fine-tuning and checkpoint selection
4. 4Negative transfer by target slice
5. 5Pretraining overlap and provenance audit

Preserve unknown labels, train clear baselines, select per-label actions and evaluate both cases and labels.

1. 1Multi-label schema and unknown targets
2. 2Binary relevance and label dependence
3. 3Per-label thresholds and action cost
4. 4Multi-label metrics and denominators
5. 5Multi-label calibration and label cardinality

Define state and reward, solve a small known model, learn from episodes and audit offline path support.

1. 1Sequential decision state and reward contract
2. 2Bellman value iteration for stock control
3. 3Episode returns, temporal difference and terminal states
4. 4Q-learning control and exploration
5. 5Offline trajectory support and simulator risk

Measure the deployment budget, compare distillation, pruning and quantization, then review target error and device cost.

1. 1Model inference budget and device profile
2. 2Teacher–student distillation objective
3. 3Structured pruning and compute shape
4. 4Affine quantization and range audit
5. 5Compression Pareto review and shadow check

Define query groups and judgments, train useful pairs, measure the top of each list, audit clicks and release within serving limits.

1. 1Ranking query groups and relevance labels
2. 2Pairwise ranking loss, ties and useful comparisons
3. 3NDCG at K with a declared query denominator
4. 4Click position bias and support for ranking evaluation
5. 5Ranking serving budget and candidate recall

Define similarity, train and mine reviewed pairs, audit exact retrieval, then measure index and serving tradeoffs.

1. 1Embedding pairs, identity labels and leakage-safe splits
2. 2Triplet margin and normalized embedding geometry
3. 3Hard-negative mining without false-negative shortcuts
4. 4Embedding retrieval recall and collapse checks
5. 5Exact versus approximate nearest-neighbor audit

Align local data contracts, aggregate compatible updates, audit participation and drift, then evaluate every deployment site.

1. 1Federated client data and target contract
2. 2Federated averaging and client weighting
3. 3Non-IID clients, local steps and update disagreement
4. 4Client participation, dropouts and stale updates
5. 5Federated evaluation by site and denominator

Audit correlated features, what-if curves, feasible edits and explanation fidelity before showing reasons to operators.

1. 1Correlated feature permutation audit
2. 2Partial dependence, ICE curves and support limits
3. 3Feasible counterfactual explanations for model decisions
4. 4Explanation stability and fidelity under model updates

Separate true events from confirmed tickets, audit selection, test prior assumptions and evaluate on adjudicated outcomes.

1. 1Positive–unlabeled target and observation state
2. 2PU label selection: SCAR and site-dependent confirmation
3. 3PU class prior and confirmation-rate sensitivity
4. 4Nonnegative PU risk objective and its assumptions
5. 5Evaluate PU models with an adjudicated sample

Define each target and mask, train separate heads, inspect loss conflicts and compare every task with a standalone baseline.

1. 1Multi-task targets and per-head label availability
2. 2Multi-task head losses with observed-label masks
3. 3Multi-task loss balance and shared-gradient conflict
4. 4Multi-task negative transfer and per-task baselines

Build unlabeled-data objectives, test transferred representations and review provenance before release.

1. 1Self-supervised objectives and valid views
2. 2Masked reconstruction targets and shortcut checks
3. 3Contrastive pairs, batch negatives and identity collisions
4. 4Representation collapse and frozen linear probes
5. 5Pretraining corpus boundaries and provenance audit

Align source spans, decode legal sequences and evaluate claims-note fields at the action level.

1. 1Sequence-label contracts and token alignment
2. 2Constrained sequence decoding and valid label paths
3. 3Span-level evaluation and extraction error taxonomy
4. 4Structured-output confidence and selective review

Define censor-aware targets, current risk sets and supported horizon decisions for operational survival models.

1. 1Time-to-event targets and censoring contracts
2. 2Survival risk sets and landmark snapshots
3. 3Time-varying survival features without future leakage
4. 4Survival-model evaluation at supported horizons
5. 5Competing events and service-policy review

Control query eligibility, batch cost, label disagreements and fixed-test evidence across acquisition rounds.

1. 1Active-learning pool eligibility and round snapshots
2. 2Active-learning batch selection under annotation cost
3. 3Active-learning annotator disagreement and adjudication
4. 4Active-learning evaluation and stopping by value

1. 1Project: evaluate a receipt-review triage model
2. 2Project: review shipment-delay and claim-risk models
3. 3Project: select a model and audit warehouse segments
4. 4Ensemble release review project
5. 5Classification and warehouse segmentation review project
6. 6Handoff triage model review project
7. 7Label budget and incremental learning project
8. 8Uncertainty and escalation review project
9. 9Distribution shift response project
10. 10Contextual bandit policy release review project
11. 11Transfer learning release review project
12. 12Multi-label parcel review project
13. 13Sequential stock-control policy review project
14. 14Model compression release project
15. 15Project: release review for maintenance-search ranking
16. 16Project: release review for visual part matching
17. 17Project: federated depot model release review
18. 18Project: release review for pump-risk explanations
19. 19Project: release review for unresolved pump failures
20. 20Project: release review for shared pump-inspection heads
21. 21Project: release review for an unlabeled parcel-image encoder
22. 22Project: release review for claims-note field extraction
23. 23Project: release review for fleet battery-replacement timing
24. 24Project: release review for invoice-label acquisition