Represent observations as tables, arrays, sequences, images, text, graphs, and signals. Identify numerical, categorical, ordinal, temporal, and missing values in real datasets.

Use Python, notebooks, NumPy, pandas, and plotting libraries to load, inspect, transform, and visualize data. Build a small reproducible analysis from a CSV file.

Work with vectors, matrices, dot products, transformations, eigenvectors, and matrix decompositions. Connect these operations to datasets, model parameters, embeddings, and dimensionality reduction.

Use conditional probability, Bayes’ rule, random variables, common distributions, expectation, and variance. Translate uncertainty in data and predictions into probability statements.

Estimate population properties from samples, quantify uncertainty, test claims, and recognize sampling bias. Apply confidence intervals, hypothesis tests, correlation, and bootstrap methods without overstating conclusions.

Use derivatives, gradients, the chain rule, and constrained optimization to describe how models improve. Implement gradient descent and diagnose learning-rate and convergence problems.

Trace the field from statistical pattern recognition and early neural networks through expert systems, kernel methods, deep learning, and foundation models. Connect changes in algorithms, data, hardware, benchmarks, and institutions to current practice.

Turn a real need into a target, unit of prediction, available features, constraints, baseline, and measurable success criterion. Check whether machine learning is appropriate before collecting data or training a model.

Plan collection, labeling, sampling, consent, documentation, retention, and access. Recognize proxy variables, historical bias, label ambiguity, leakage risks, and populations missing from the dataset.

Summarize distributions, relationships, missingness, class balance, outliers, and suspicious patterns with visual and numerical checks. Produce an exploratory report that guides later modeling choices.

Clean inconsistent values, encode categories, scale measurements, handle missing data, and split data safely. Build reusable preprocessing pipelines that apply identical transformations during training and inference.

Create useful predictors from domain signals while avoiding target leakage. Work with interactions, bins, aggregations, dates, text counts, and learned representations, then assess whether each feature adds value.

Separate training, validation, and test data; choose metrics that match the real cost of errors; and compare against meaningful baselines. Use cross-validation, confidence intervals, and error analysis to judge whether gains are reliable.

Study generalization, overfitting, underfitting, bias, variance, regularization, and the effect of dataset size. Use learning curves and validation results to decide whether to change the data, features, or model.

Fit continuous outcomes with least squares, regularization, polynomial terms, and robust losses. Check residuals and assumptions, quantify uncertainty, and interpret coefficients carefully.

Model counts, probabilities, and other non-Gaussian outcomes with logistic, Poisson, and related generalized linear models. Produce calibrated class probabilities and explain how features affect predictions.

Make predictions from nearby examples using distance metrics and local weighting. Address scaling, high-dimensional distance, computational cost, and the choice of neighborhood size.

Apply Bayes’ rule to classification with Naive Bayes and related probabilistic models. Estimate class probabilities, handle sparse features, and recognize when independence assumptions help or fail.

Build decision trees, control their complexity, and interpret their splits. Use random forests, gradient boosting, and stacking to improve predictive performance while managing variance and calibration.

Find separating boundaries with margins, kernels, and support vectors. Tune soft-margin and kernel parameters, then compare support vector machines with linear and tree-based alternatives.

Group unlabeled observations using k-means, hierarchical clustering, density-based methods, and mixture models. Choose distance measures and validate whether clusters are stable and useful rather than merely convenient.

Compress and visualize high-dimensional data with principal component analysis, matrix factorization, manifold methods, and modern neighborhood embeddings. Distinguish techniques suited to denoising, visualization, and downstream modeling.

Estimate data density and detect rare or unusual observations with statistical, distance-based, isolation, and reconstruction methods. Set alert thresholds and evaluate results when confirmed anomaly labels are scarce.

Represent dependencies among uncertain variables with Bayesian networks, Markov models, and factor graphs. Perform inference, estimate hidden states, and use probabilistic structure when uncertainty must be modeled directly.

Search model settings with grids, random search, Bayesian optimization, and resource-aware methods. Design tuning experiments that avoid test-set leakage and wasteful computation.

Train models when labeled examples are limited by combining labeled and unlabeled data. Use pseudo-labeling, consistency objectives, active learning, and human review to spend labeling effort where it matters most.

Build neural networks from layers, activations, losses, backpropagation, initialization, and regularization. Train a multilayer network and diagnose unstable gradients, poor convergence, and overfitting.

Learn compact representations with embeddings, metric learning, contrastive objectives, and self-supervised pretraining. Evaluate whether a representation transfers to retrieval, clustering, or prediction tasks.

Process images with convolutions, pooling, residual connections, augmentation, and transfer learning. Build systems for classification, detection, and segmentation while checking robustness across visual conditions.

Model ordered data with recurrent networks, gated units, temporal convolutions, and sequence-to-sequence architectures. Handle variable-length inputs, masking, teacher forcing, and long-range dependencies.

Use attention, positional information, encoder-decoder structures, and transformer training objectives. Implement a small transformer and compare its computational and modeling tradeoffs with recurrent networks.

Represent entities and relationships with node embeddings, message passing, and graph neural networks. Apply graph learning to node classification, link prediction, and whole-graph prediction.

Model trend, seasonality, autocorrelation, interventions, and uncertainty in time-indexed data. Build backtests for statistical, tree-based, and neural forecasting methods without leaking future information.

Build personalized ranking systems with neighborhood methods, matrix factorization, learning-to-rank, and two-stage retrieval and ranking. Address cold starts, feedback loops, diversity, and offline versus online evaluation.

Separate association from intervention using causal graphs, potential outcomes, randomized experiments, matching, weighting, and natural experiments. Estimate treatment effects while stating the assumptions that make a causal claim possible.

Model agents that act over time using states, actions, rewards, policies, and value functions. Implement bandits, temporal-difference learning, policy gradients, and actor-critic methods while addressing exploration and unsafe behavior.

Adapt models to changing streams with incremental updates, sliding windows, and drift detection. Decide when to retrain, recalibrate, or retire a model as real-world behavior changes.

Generate and reconstruct data with autoencoders and variational autoencoders. Work with latent spaces, reconstruction objectives, sampling, and the tradeoff between fidelity and useful structure.

Train generators and discriminators in adversarial competition. Address unstable training, mode collapse, evaluation, and practical uses such as image synthesis and domain adaptation.

Generate images and other data by learning to reverse a gradual noising process. Cover denoising objectives, latent diffusion, conditioning, guidance, sampling speed, and evaluation of generated outputs.

Study large pretrained models built from broad datasets, including scaling behavior, tokenization, pretraining objectives, instruction tuning, and emergent capabilities. Compare language, vision, audio, and multimodal foundation-model designs.

Design reliable language-model interactions with clear instructions, examples, structured outputs, tool schemas, and automated checks. Evaluate prompts as versioned system components rather than one-off text snippets.

Connect language models to external knowledge through document preparation, embeddings, vector search, reranking, context assembly, and citation checks. Build and evaluate a retrieval-augmented generation pipeline for answer quality and grounding.

Adapt pretrained models with supervised fine-tuning, parameter-efficient methods, preference optimization, and distillation. Prepare training data, select checkpoints, measure regressions, and control compute and memory costs.

Combine text, images, audio, video, and sensor data through shared representations and cross-modal attention. Build systems for captioning, visual question answering, speech-language tasks, and multimodal retrieval.

Build language-model systems that plan, call tools, maintain state, and complete multi-step tasks. Set execution limits, validate tool outputs, recover from failures, and measure whether agentic designs outperform simpler workflows.

Create realistic artificial examples for augmentation, simulation, testing, and privacy-sensitive development. Measure utility, detect memorization, and avoid reinforcing biases or contaminating evaluation sets.

Automate preprocessing, model selection, feature search, and tuning with AutoML systems. Use automation as a controlled search process while retaining human responsibility for framing, validation, and deployment decisions.

Organize reliable datasets with batch and streaming pipelines, validation checks, lineage, versioning, and feature stores. Prevent training-serving skew and make data failures visible before they affect predictions.

Use experiment trackers, configuration files, dataset and model versioning, random seeds, and reproducible environments. Turn notebook experiments into results another practitioner can rerun and audit.

Accelerate training with GPUs, mixed precision, distributed data parallelism, model parallelism, and efficient input pipelines. Estimate memory, throughput, communication, energy, and cost tradeoffs before scaling a workload.

Package models behind batch jobs, APIs, streaming services, or on-device runtimes. Address latency, throughput, serialization, dependency management, fallback behavior, and safe rollout strategies.

Connect data validation, training, testing, model registration, deployment, and rollback through automated pipelines. Apply containers, CI/CD, orchestration, infrastructure controls, and approval gates to production ML systems.

Track input quality, drift, latency, cost, prediction quality, calibration, and downstream outcomes after release. Design alerts, feedback loops, retraining triggers, incident response, and model retirement procedures.

Compress and optimize models through quantization, pruning, distillation, and hardware-aware compilation. Deploy inference to phones, browsers, sensors, and embedded devices under tight power, memory, privacy, and latency limits.

Train across decentralized data sources with federated optimization, secure aggregation, and privacy accounting. Apply differential privacy and related protections while measuring their effect on model quality and equity.

Explain individual predictions and global model behavior with coefficients, feature importance, partial dependence, local attribution, examples, and counterfactuals. Match each explanation method to its assumptions, audience, and failure modes.

Measure unequal errors, representation gaps, accessibility harms, and feedback loops across affected groups. Compare fairness criteria, document tradeoffs, and involve stakeholders in deciding what acceptable performance means.

Threat-model poisoning, evasion, model extraction, prompt injection, data leakage, insecure artifacts, and dependency compromise. Add access controls, adversarial testing, provenance checks, sandboxing, and incident procedures across the ML lifecycle.

Set ownership, approval, documentation, audit, risk classification, human oversight, and change-control rules for ML systems. Use model cards, data documentation, impact assessments, evaluation records, and applicable legal or sector requirements.

Carry a real problem from stakeholder request through data collection, baseline construction, model training, error analysis, deployment, monitoring, and follow-up. Produce a versioned repository, model documentation, service interface, and operational handoff.

Choose among research, data, modeling, ML engineering, product, safety, and governance roles based on the work each actually performs. Build proof-of-skill projects, assess relevant certifications, contribute to open work, read papers critically, and maintain a reliable practice for staying current.