Artificial Intelligence

Set up Python, notebooks, packages, files, and simple scripts used in AI work. You will run small experiments, save results, and read error messages without treating the computer as a black box.

Work with tables, text, images, audio, labels, features, training data, and test data. You will inspect a small dataset, find missing values, and prepare it for a first model.

Use vectors, matrices, distances, dot products, and simple transformations as the language of many AI models. The chapter keeps the math visual and tied to examples like similarity search and image pixels.

Use probability, distributions, averages, variance, sampling, correlation, and uncertainty in AI decisions. You will calculate simple risks and see why models can be confident and still be wrong.

Use functions, slopes, gradients, loss, and iterative improvement to see how models get trained. You will fit a tiny model by repeatedly reducing error.

Trace AI from early logic and cybernetics through expert systems, statistical machine learning, deep learning, and foundation models. The chapter shows how funding, data, hardware, and expectations shaped what AI practitioners do today.

Turn a vague request into an AI task with inputs, outputs, constraints, risks, and a measurable target. You will choose success metrics that match real use, not just impressive demos.

Follow a small AI project from question to dataset, baseline model, evaluation, error analysis, and a short report. This gives you a complete loop before the course adds more powerful techniques.

Represent problems as states, actions, paths, and goals. You will solve toy route, puzzle, and scheduling problems with breadth-first search, depth-first search, and cost-based search.

Use heuristics to make search faster when brute force is too slow. You will apply greedy search and A* search while judging when a shortcut is helpful or misleading.

Model competitive settings where another player or system pushes back. You will build game trees, use minimax, add alpha-beta pruning, and evaluate simple board-game positions.

Solve problems where choices must satisfy rules, such as timetables, maps, Sudoku, and resource assignments. You will use variables, domains, constraints, backtracking, and constraint propagation.

Represent facts, categories, relationships, and meanings so a system can reason with them. You will build small knowledge graphs and ontologies for a practical domain.

Use formal logic to make reliable inferences from stated facts and rules. You will work with propositions, predicates, unification, and simple automated proofs.

Create action sequences for agents that need to reach goals under constraints. You will model preconditions, effects, plans, schedules, and failures in small task domains.

Build rule-based systems that encode expert decisions in if-then form. You will see where expert systems still help, where they become brittle, and how rule maintenance works.

Reason when facts are incomplete, noisy, or changing. You will update beliefs with evidence and compare choices under uncertainty.

Model connected uncertain events with Bayesian networks and related graphical models. You will perform simple inference for diagnosis, prediction, and explanation.

Connect probabilities to actions through utility, cost, risk, and expected value. You will build decision trees and choose actions when outcomes are uncertain.

Run the standard machine learning cycle: define the target, split data, train a baseline, measure performance, and improve carefully. You will also track assumptions that can make a model fail outside the training data.

Predict numbers with linear regression and related models. You will fit, interpret, regularize, and evaluate regression models on real tabular data.

Predict categories with logistic regression, trees, forests, boosting, and support vector machines. You will compare decision boundaries, probabilities, and error types for real classification tasks.

Transform raw data into useful model inputs without leaking answers from the future or the test set. You will encode categories, scale numbers, handle missing data, and build repeatable preprocessing pipelines.

Measure whether a model is truly useful with train-test splits, cross-validation, baselines, calibration, and error analysis. You will choose metrics for imbalanced data, ranking, regression, and classification.

Find structure in unlabeled data with clustering, dimensionality reduction, and representation learning. You will use these methods for customer grouping, visualization, and dataset inspection.

Detect rare, unusual, or suspicious patterns in data. You will compare statistical thresholds, isolation forests, reconstruction error, and practical review workflows.

Model data ordered through time, such as sales, traffic, sensors, or demand. You will handle trends, seasonality, lagged features, forecasting horizons, and backtesting.

Recommend items, content, or people using collaborative filtering, content-based methods, ranking models, and feedback loops. You will build a simple recommender and check for cold-start and popularity bias problems.

Separate correlation from cause when decisions can change outcomes. You will use causal graphs, confounding, experiments, natural experiments, and uplift thinking for practical AI decisions.

Train agents that learn from rewards through trial and error. You will use Markov decision processes, value functions, policy learning, exploration, and simple simulated environments.

Build neural networks from neurons, layers, activations, weights, loss functions, and backpropagation. You will train a small network and inspect what changes during training.

Train deep models with batches, optimizers, initialization, normalization, regularization, and learning-rate schedules. You will diagnose overfitting, underfitting, unstable training, and slow training.

Use GPUs, TPUs, memory, batching, and parallelism as practical limits on modern AI work. You will estimate compute needs and avoid common hardware bottlenecks.

Process images with convolutional neural networks and core computer vision tasks. You will build models for image classification, object detection, segmentation, and visual error analysis.

Model ordered data with recurrent networks, embeddings, attention ideas, and encoder-decoder systems before transformers took over. This chapter prepares you to see what transformers changed and what older sequence methods still do well.

Process human language with tokenization, text classification, information extraction, semantic similarity, and evaluation. You will build NLP pipelines that work before adding large language models.

Work with waveforms, spectrograms, speech recognition, speaker features, and audio classification. You will build a small audio model and judge errors that text-only AI would miss.

Represent networks of people, products, molecules, citations, or web pages with nodes and edges. You will use graph features and graph neural networks for prediction on connected data.

Connect perception, planning, control, and learning in machines that act in the physical world. You will reason about sensors, actuators, maps, motion, simulation, and safety margins.

Use self-attention, positional information, token embeddings, and transformer blocks. You will trace how a transformer processes a sequence and why the architecture scales so well.

Work with large pretrained models that can handle language and other tasks through broad training. You will cover pretraining, scaling, instruction tuning, context windows, emergent behavior, and common failure modes.

Shape model behavior with clear prompts, examples, roles, constraints, and structured outputs. You will create reusable prompt patterns and test them against messy user inputs.

Connect generative models to external documents, databases, and search results. You will build a retrieval-augmented generation pipeline with chunking, embeddings, vector search, reranking, citations, and answer checks.

Adapt pretrained models to specific tasks or domains with fine-tuning, transfer learning, adapters, and LoRA-style methods. You will compare when fine-tuning beats prompting or retrieval and how to prepare training data.

Evaluate generative AI with task tests, human review, automated checks, red teaming, and safety filters. You will design guardrails for hallucinations, toxic output, privacy leaks, and instruction conflicts.

Build systems where models call tools, use memory, plan steps, and act across software environments. You will design agent loops, tool schemas, permissions, stop conditions, and failure recovery.

Generate images and other media with denoising diffusion, latent spaces, conditioning, guidance, and sampling. You will compare prompt control, image-to-image editing, inpainting, and practical evaluation of generated media.

Combine text, images, audio, video, and structured data in one AI system. You will work with cross-modal embeddings, vision-language models, document understanding, and multimodal evaluation.

Improve AI systems by improving data quality, coverage, labels, and feedback loops. You will use dataset audits, active learning, weak supervision, synthetic data, and human review queues.

Use AI tools to write, review, test, document, and refactor code while keeping human responsibility clear. You will practice safe workflows for code generation, debugging, dependency risk, and verification.

Design AI that is useful to people in real settings, not just accurate in a notebook. You will shape interfaces, feedback, confidence displays, handoffs, and human override paths.

Choose the pieces of an AI system: data sources, batch jobs, streaming, model APIs, storage, retrieval, application logic, and user interfaces. You will sketch architectures for predictive, generative, and agentic products.

Package models for real use with APIs, containers, batch inference, real-time serving, and scalable endpoints. You will compare latency, cost, reliability, and update strategies.

Track experiments, datasets, model versions, parameters, and reproducible runs. You will use MLOps practices to move from isolated notebooks to team-ready AI work.

Watch deployed models for drift, data quality issues, performance drops, cost spikes, and user harm. You will set alerts, investigate incidents, and plan rollback or retraining.

Make models smaller, faster, and cheaper with quantization, distillation, pruning, caching, and edge deployment. You will decide when to run AI on a server, browser, phone, or embedded device.

Train or use models when data is sensitive or spread across organizations. You will apply privacy risk thinking, differential privacy, secure aggregation, and federated learning patterns.

Protect AI systems from evasion, poisoning, prompt injection, data exfiltration, model theft, and unsafe tool calls. You will threat-model both classical ML systems and LLM applications.

Explain model behavior with interpretable models, feature importance, counterfactuals, saliency maps, and example-based explanations. You will match explanation methods to users, risks, and evidence needs.

Find and reduce unfair performance across groups, contexts, and access levels. You will audit datasets, compare fairness metrics, document tradeoffs, and design mitigation steps.

Handle risks from powerful AI systems, including misalignment, misuse, loss of control, and high-stakes deployment. You will work with safety cases, red-team findings, capability limits, and staged release decisions.

Apply laws, policies, standards, documentation duties, and responsible release practices that shape AI work. You will create model cards, data documentation, risk registers, and approval checklists for realistic projects.

Deliver a complete AI product from stakeholder request through data work, model choice, evaluation, deployment plan, monitoring plan, and handoff. The capstone forces tradeoffs among accuracy, cost, safety, privacy, and user value.

Map AI roles, specialties, portfolios, interviews, open-source paths, research paths, certifications, and ways to stay current. You will leave with concrete proof-of-skill artifacts and a plan for continued growth.