Introduction to Artificial Intelligence

Course Description

CAI4002 – Introduction to Artificial Intelligence is a 3-credit upper-division course introducing the concepts, techniques, and tools of artificial intelligence: what AI systems do, the learning paradigms behind them, how they are evaluated, and the social and ethical questions their deployment raises. Some Florida institutions title it "Foundations of AI."

The CAI prefix denotes computing and artificial intelligence, and this is a computer science course rather than a survey for general audiences — prerequisites typically include data structures and statistics, and the material assumes programming fluency and comfort with mathematical notation.

Content covers the history and goals of AI — symbolic approaches, the AI winters, and the shift to statistical and learned methods; intelligent agents — environments, rationality, and agent architectures; search — uninformed and informed search, heuristics, A*, and local search; adversarial search — game playing, minimax, and alpha-beta pruning; constraint satisfaction; knowledge representation and reasoning — logic, inference, and ontologies; uncertainty — probability, Bayes' rule, and Bayesian networks; machine learning paradigms — supervised, unsupervised, and reinforcement learning; supervised learning — classification and regression, training and test splits, overfitting, and evaluation metrics; neural networks — perceptrons, backpropagation, and deep architectures; natural language processing — representation, embeddings, and language models; computer vision at an introductory level; reinforcement learning — agents, rewards, and policy; tools and frameworks; AI careers and the shape of the field; and ethics and social impact — bias, fairness, transparency, accountability, privacy, labor effects, and safety.

Offered at Florida institutions with computing programs.

Learning Outcomes

Required Outcomes

• Describe the history of artificial intelligence and its major paradigm shifts.
• Define an intelligent agent and characterize task environments.
• Formulate a problem as a search problem with states, actions, and goals.
• Implement and compare uninformed search strategies.
• Design admissible heuristics and apply informed search including A*.
• Apply local search and optimization methods.
• Apply minimax and alpha-beta pruning to adversarial problems.
• Formulate and solve constraint satisfaction problems.
• Represent knowledge using logic and perform inference.
• Apply probability and Bayes' rule to reasoning under uncertainty.
• Describe Bayesian networks and their use in inference.
• Distinguish supervised, unsupervised, and reinforcement learning.
• Train and evaluate a supervised model on a dataset.
• Apply training, validation, and test splits correctly and explain why.
• Diagnose overfitting and underfitting and apply appropriate remedies.
• Select and interpret evaluation metrics appropriate to a task.
• Describe neural network architecture and the role of backpropagation.
• Describe natural language processing including embeddings and language models.
• Describe computer vision tasks and approaches at an introductory level.
• Describe reinforcement learning including states, actions, rewards, and policy.
• Use standard tools and frameworks to build and evaluate a model.
• Identify sources of bias in data and models and describe mitigation approaches.
• Evaluate an AI application for fairness, transparency, and accountability.
• Describe privacy, labor, and safety implications of AI deployment.

Optional Outcomes

• Complete an applied AI project end to end.
• Describe large language models and their capabilities and limitations.
• Describe AI regulation and emerging governance frameworks.
• Describe explainability and interpretability methods.
• Describe AI applications in a specific domain such as health care or robotics.
• Read and present a research paper.

Major Topics

Required Topics

• What AI is — definitions, goals, and the moving target problem.
• History — symbolic AI, expert systems, AI winters, and the statistical turn.
• Intelligent agents — rationality, environments, and architectures.
• Problem formulation — states, actions, transitions, and goals.
• Uninformed search — breadth-first, depth-first, and uniform cost.
• Informed search — heuristics, greedy search, and A*.
• Local search — hill climbing, simulated annealing, and genetic algorithms.
• Adversarial search — minimax and alpha-beta pruning.
• Constraint satisfaction — formulation and solving techniques.
• Knowledge representation — propositional and first-order logic.
• Inference — resolution, forward and backward chaining.
• Uncertainty — probability, independence, and Bayes' rule.
• Bayesian networks — structure and inference.
• Learning paradigms — supervised, unsupervised, reinforcement.
• Supervised learning — classification and regression methods.
• Model evaluation — splits, cross-validation, and metrics.
• Overfitting — bias-variance, regularization, and diagnosis.
• Unsupervised learning — clustering and dimensionality reduction.
• Neural networks — perceptrons, layers, and backpropagation.
• Deep learning — architectures and what changed.
• Natural language processing — representation, embeddings, and models.
• Computer vision — tasks and convolutional approaches.
• Reinforcement learning — agents, rewards, and policies.
• Tools — Python, scikit-learn, and deep learning frameworks.
• Bias and fairness — sources, measurement, and mitigation.
• Ethics and society — transparency, accountability, privacy, labor, safety.

Optional Topics

• Applied end-to-end project.
• Large language models and their limitations.
• AI regulation and governance.
• Explainability and interpretability.
• Domain applications.
• Research paper reading and presentation.

Resources & Tools

• Artificial Intelligence: A Modern Approach (Russell & Norvig), Pearson — the field's dominant textbook by a wide margin; assume you will use it and keep it.
• Hands-On Machine Learning with Scikit-Learn, Keras, and TensorFlow (Géron) — the best practical companion; excellent for the learning half of the course.
• Python with NumPy, pandas, and scikit-learn — free; the standard toolchain, and what employers expect.
• Google Colab — free notebooks with GPU access; removes the hardware barrier entirely for coursework.
• Kaggle — free datasets, notebooks, and competitions; the best available place to practice on real data and see how others approached the same problem.
• Hugging Face — free models, datasets, and tutorials; the practical entry point to modern NLP.
• fast.ai and Andrew Ng's courses — free or low cost; strong parallel tracks if the course moves fast.
• arXiv — free preprints; the field publishes here first, which is worth knowing.
• UCI Machine Learning Repository and Papers with Code — free datasets and reproducible implementations.

Career Pathways

• Machine Learning Engineer — the most common destination; software engineering plus modeling.
• Data Scientist (SOC 15-2051) — analysis, modeling, and communication.
• Software Developer (SOC 15-1252) — increasingly with AI components in ordinary products.
• Data Engineer — pipelines and infrastructure; in higher demand than most students realize.
• AI/ML Operations Engineer — deployment, monitoring, and reliability.
• Computer Vision or NLP Engineer — specialization with experience.
• Research Scientist — typically requires a graduate degree.
• AI Governance, Policy, or Audit — an emerging area combining technical and regulatory literacy.
• Domain analyst roles — health care, finance, logistics, and defense, where domain knowledge plus modeling is scarce.

Florida's demand concentrates in Tampa, Orlando, Miami, and Jacksonville, with substantial pull from the defense, simulation, and training sector in Central Florida, the aerospace industry on the Space Coast, health systems, financial services, and logistics. Much of this work is also remote-eligible, which widens the market and the competition alike — portfolios matter more than transcripts.

Special Information

⚠ The prerequisites are real: this is a computer science course

Worth confirming before registering. Florida institutions commonly require COP3530 data structures and STA2023 statistics, and the course assumes fluent programming, comfort with algorithmic complexity, and enough mathematics — linear algebra notation, probability, and calculus concepts — to follow derivations without being taught them.

A student without that background can still learn a great deal about AI from the many excellent general resources available, but will struggle in this course specifically. The most common preparation gap is probability: the learning half of the course rests on it continuously, and a student shaky on conditional probability and Bayes' rule should review before the term rather than during it.

The field moved, and the textbook is partly historical

Useful framing for a course that spans two eras. Roughly the first half of a traditional AI course — search, logic, constraint satisfaction, planning — is symbolic AI, which dominated the field for decades and remains genuinely important in scheduling, verification, routing, and game playing. The second half is statistical and learned methods, which is where nearly all recent progress and nearly all industry work sits.

Students sometimes conclude the first half is obsolete. It is not — A* is running in every navigation app, constraint solvers schedule real operations, and the search framing teaches how to formalize a problem, which is a transferable skill. But it is worth knowing which era you are in at any point in the course, and that the practical job market is overwhelmingly on the learning side.

⚠ Evaluation is where beginners fool themselves

The most consequential technical discipline in the course, and the error that survives into professional work. A model that performs beautifully on the data it was trained on has demonstrated nothing; the entire question is how it performs on data it has not seen.

The specific failure modes worth internalizing: training on the test set, including indirectly by tuning hyperparameters against it — which is why a separate validation split exists; data leakage, where information unavailable at prediction time creeps into the features; accuracy on imbalanced data, where predicting the majority class always scores 99 percent and is useless, which is why precision, recall, and the confusion matrix exist; and distribution shift, where the model is deployed on data that does not resemble what it was trained on. A student who can explain why their model's reported number is trustworthy is doing the actual work.

⚠ Bias and ethics are technical content, not an appended lecture

Worth treating seriously rather than as the week before finals. AI systems learn from data, data reflects the world including its inequities, and a model trained on historical decisions reproduces the pattern of those decisions. Documented cases span hiring, lending, medical triage, and criminal risk assessment — and in several, the harm was invisible to the developers because the aggregate metrics looked fine.

The technically substantive points: bias enters through data collection, labeling, feature selection, and objective choice, not merely through prejudiced intent; fairness has multiple formal definitions that are mathematically incompatible, so "make it fair" is not a specification; removing a protected attribute does not remove its influence, because correlated features proxy for it; and evaluation should be disaggregated by subgroup, because aggregate accuracy hides subgroup failure. These are engineering considerations, and a graduate who can raise them credibly is more valuable, not less employable.

Build things, and be able to explain the ones you built

The most actionable career advice. AI hiring runs substantially on demonstrated work, and the entry cost is essentially zero: Google Colab provides free GPU notebooks, Kaggle provides real datasets and public solutions to compare against, and Hugging Face provides pretrained models.

Two qualifications on the standard advice. First, a portfolio of tutorial reproductions signals very little, because everyone has them; a project on a dataset you found or assembled, addressing a question you chose, signals a great deal. Second, be able to explain your own choices — why that model, why that metric, what you tried that failed, what the limitations are. Interviews probe exactly there, and a candidate who cannot defend a design decision on their own project is transparent.

Learn the tooling that surrounds the model, not just the model

A gap between coursework and industry worth naming. Courses focus on model selection and training. Most professional machine learning work is data — acquiring it, cleaning it, labeling it, building pipelines, and monitoring what happens after deployment when the data shifts. Job titles reflect this: data engineering and ML operations roles are frequently easier to enter and are in higher demand than research-adjacent modeling roles.

Practically, a student who adds SQL, version control, basic cloud deployment, and the discipline of reproducible notebooks to this coursework is substantially more employable than one who can only train a model in a notebook that runs once. The pairing with CTS2375C cloud computing or a database course is a strong combination in the Florida market.

Numbering and program context

The CAI prefix denotes computing and artificial intelligence in Florida's numbering system, and titles vary — "Introduction to Artificial Intelligence" and "Foundations of AI" both appear for CAI4002. Prerequisites differ by institution; Florida Gulf Coast lists COP3530 and STA2023. Related coursework sits under COP (programming and data structures), CAP (applications including machine learning and robotics), CDA (architecture), and STA (statistics), and Florida institutions have added AI-specific degrees and certificates rapidly in recent years, including A.S. and College Credit Certificate options. SCNS equivalency applies to the same number at the same level, never across numbers; because this is a 4000-level major course, confirm that a receiving institution accepts it toward the major rather than as elective credit.