Course Description
EEE 4773 Fundamentals of Machine Learning is the electrical engineering department’s course in learning from data. Its distinguishing feature, compared with a computer science treatment of the same subject, is that it is built on signals and probability rather than on algorithms and data structures — the prerequisite is a signals and systems course, and the material is presented as inference about noisy measurements.
The Statewide Course Numbering System describes it as an “overview of machine intelligence and the role of machine learning in a variety of real-world problems. Probability and statistics to handle uncertain data. Topics covered include: learning models from data in both a supervised and unsupervised fashion, linear models and nonlinear models for classification, and linear dimensionality reduction.” The statewide prerequisites are EEL 3135 or equivalent and STA 3032 or equivalent.
Three Florida institutions carry it, all at 3 credits: the University of Florida as Fundamentals of Machine Learning (matching the statewide title), and Florida A&M University and Florida State University as Machine Learning.
⚠⚠ Two warnings before anything else, and both matter for registration.
First: the University of Florida runs two machine learning courses under adjacent numbers, and they are not the same course. EEE 3773 Introduction to Machine Learning is a 4-credit course requiring only EEL 3135. EEE 4773 Fundamentals of Machine Learning is a 3-credit course requiring EEL 3135 and EEL 3850 (probability), each with a minimum grade of C. Confusing them is easy and consequential — different credit values, different prerequisites, different levels.
Second: FAMU and FSU teach a broader subject than the statewide description. The FAMU–FSU College of Engineering describes its course as introducing “theory and engineering applications of machine learning including neural networks, fuzzy logic, genetic algorithms, and supervised/unsupervised learning algorithms.” Fuzzy logic and genetic algorithms are computational intelligence topics that sit outside the statistical machine learning the statewide description sets out. This is a genuine difference in emphasis, and the Special Information section says what follows from it.
Learning Outcomes
Required Outcomes
- Formulate a real-world problem as a machine learning task, and identify whether it is supervised, unsupervised, a regression or a classification problem.
- Apply probability and statistics to reason about uncertain data: distributions, conditional probability, Bayes’ rule, expectation and variance, and maximum likelihood estimation.
- Derive and apply linear regression, including the least-squares solution in matrix form and its probabilistic interpretation.
- Derive and apply linear classifiers — the perceptron, logistic regression, linear discriminant analysis — and state their decision boundaries.
- Explain the bias–variance trade-off and relate it to model capacity, training set size and generalisation error.
- Apply regularisation (ridge, lasso) and explain its effect on model complexity and on feature selection.
- Design and interpret a validation methodology: train/validation/test splitting, cross-validation, and the reasons a test set must be held out.
- Select and interpret evaluation metrics appropriate to the problem — accuracy, precision and recall, F1, ROC and AUC, confusion matrices — and explain why accuracy misleads on imbalanced data.
- Apply non-linear classification methods: k-nearest neighbours, decision trees, support vector machines with kernels, and feedforward neural networks.
- Apply Bayesian classification, including the naive Bayes assumption and when it is reasonable.
- Apply unsupervised methods: k-means and hierarchical clustering, and Gaussian mixture models with expectation–maximisation.
- Apply linear dimensionality reduction — principal component analysis, and linear discriminant analysis in its projection role — and interpret the resulting components.
- Implement, train and evaluate a complete machine learning pipeline in MATLAB or Python on a real dataset.
Optional Outcomes
- Apply ensemble methods: bagging, random forests and boosting.
- Train deeper neural networks and apply convolutional architectures.
- Apply non-linear dimensionality reduction and manifold learning.
- Apply reinforcement learning fundamentals.
- Apply fuzzy logic systems and fuzzy inference (FAMU/FSU).
- Apply genetic algorithms and evolutionary optimisation (FAMU/FSU).
- Apply machine learning to signal-specific problems — time series, speech, images — using domain feature extraction.
- Analyse fairness, bias and interpretability in deployed models.
Major Topics
Required Topics
- Machine intelligence overview — the landscape of learning problems, where machine learning succeeds and where it does not, and the relationship between machine learning, statistics and signal processing.
- Probability and statistics for uncertain data — random variables and distributions, joint and conditional probability, Bayes’ rule, expectation and covariance, the multivariate Gaussian, and maximum likelihood and MAP estimation.
- Linear regression — the least-squares formulation, the normal equations and their matrix solution, the Gaussian-noise probabilistic interpretation, polynomial and basis-function expansion, and overfitting as capacity increases.
- Linear models for classification — the perceptron and its convergence, logistic regression and the sigmoid, the cross-entropy objective, gradient descent, linear discriminant analysis, and multi-class extension.
- Generalisation — training and test error, the bias–variance decomposition, model capacity, the learning curve, and why more data and more capacity trade against one another.
- Regularisation — ridge and lasso penalties, their Bayesian interpretation as priors, the sparsity property of the lasso, and hyperparameter selection.
- Model selection and validation — the train/validation/test discipline, k-fold and leave-one-out cross-validation, nested cross-validation, and the ways information leaks between splits.
- Evaluation — confusion matrices, precision and recall, F-measure, ROC curves and AUC, class imbalance and the failure of accuracy, and the asymmetry of error costs.
- Non-linear methods — k-nearest neighbours and the curse of dimensionality; decision trees, splitting criteria and pruning; support vector machines, margins, the dual formulation and kernels; feedforward neural networks and backpropagation.
- Bayesian classification — the Bayes optimal classifier, discriminant functions, naive Bayes and the independence assumption, and Gaussian class-conditional models.
- Unsupervised learning — k-means and its initialisation sensitivity, hierarchical clustering and linkage criteria, Gaussian mixture models and expectation–maximisation, and cluster validity.
- Linear dimensionality reduction — principal component analysis derived from variance maximisation and from reconstruction error, the eigenvalue problem and the singular value decomposition, component selection, whitening, and LDA as supervised projection.
- Practice — feature scaling and normalisation, handling missing data, the pipeline from raw data to evaluated model, and reproducibility.
Optional Topics
- Ensembles: bagging, random forests, AdaBoost and gradient boosting.
- Deep learning: convolutional networks, regularisation by dropout, and transfer learning.
- Non-linear dimensionality reduction: kernel PCA, t-SNE, UMAP.
- Reinforcement learning: Markov decision processes, value iteration, Q-learning.
- Fuzzy logic — fuzzy sets, membership functions, fuzzy inference and control (FAMU/FSU).
- Genetic and evolutionary algorithms — encoding, selection, crossover, mutation, and application to optimisation (FAMU/FSU).
- Time-series and sequence models; feature extraction from signals and images.
- Fairness, bias, interpretability and explainability.
Resources & Tools
- Pattern Recognition and Machine Learning (Christopher Bishop) is the reference standard and is written from exactly the probabilistic standpoint this course takes; The Elements of Statistical Learning (Hastie, Tibshirani and Friedman) is the statistical reference and is free from the authors. An Introduction to Statistical Learning (James, Witten, Hastie and Tibshirani), also free, is the most accessible and is an excellent first text.
- Machine Learning: A Probabilistic Perspective (Murphy) and Pattern Classification (Duda, Hart and Stork) are the other common adoptions; Duda and Hart in particular is the classic in the engineering tradition this course belongs to.
- Python with scikit-learn, NumPy, pandas and Matplotlib is the dominant environment and is what industry uses; PyTorch or TensorFlow where neural networks are taken further. MATLAB with the Statistics and Machine Learning Toolbox is common in engineering departments and is what FAMU/FSU name explicitly.
- Datasets — the UCI Machine Learning Repository, Kaggle, and OpenML for coursework and projects. ⚠ Working with a real, messy dataset rather than a cleaned teaching one is where most of the learning happens.
- Environments — Jupyter notebooks and Google Colab, the latter providing free GPU access, which matters if the course reaches neural networks.
- Reference material — the IEEE Signal Processing Society and its Transactions on Signal Processing and Transactions on Pattern Analysis and Machine Intelligence; NeurIPS and ICML proceedings, which are freely available and are where the field publishes.
Career Pathways
- Machine learning engineer (SOC 15-2051, Data Scientists; SOC 15-1252, Software Developers) — the direct destination, and a role in which an electrical engineering background is a genuine advantage rather than a handicap, because so many applications involve sensor and signal data.
- Data scientist (SOC 15-2051) and data analyst.
- Signal processing engineer with machine learning skills — arguably the strongest positioning available from this course, since the combination is scarcer than either alone.
- Computer vision engineer and speech and audio engineer.
- Robotics and autonomy engineer — perception and control, a substantial employment area in Florida’s defence sector.
- Research engineer in industrial or national laboratories; and this course is a common route into graduate study.
- Florida employers include L3Harris (Melbourne and Palm Bay — automatic target recognition, signal classification and autonomy), Lockheed Martin (Orlando — AI for simulation and training, and autonomous systems), Northrop Grumman (Melbourne), Leidos and SAIC, the Orlando simulation and training cluster, NextEra Energy and Florida Power & Light (grid forecasting and predictive maintenance), AdventHealth, Moffitt Cancer Center and Mayo Clinic (Jacksonville) for clinical machine learning, and the substantial financial-services presence in South Florida.
- ⚠ A realistic note. Machine learning is a competitive field with a large supply of applicants, many with computer science or statistics graduate degrees. An electrical engineering graduate competes best where the data are signals — radar, audio, biomedical, sensor, RF — because domain understanding is what a generic machine learning candidate lacks. Pair this course with DSP (EEE 4510) and lean on that combination.
Special Information
⚠⚠ At UF, do not confuse EEE 3773 with EEE 4773
| EEE 3773 | EEE 4773 |
|---|
| Title | Introduction to Machine Learning | Fundamentals of Machine Learning |
| Credits | 4 | 3 |
| Prerequisites | EEL 3135 | EEL 3135 and EEL 3850 (probability), each C or better |
| Level | Junior | Senior |
These are two courses, not two names for one. The credit difference alone affects a degree audit, and the probability prerequisite on EEE 4773 signals a more statistically rigorous treatment. Check your degree requirements against the number, not the title, and confirm with an advisor which your programme expects.
⚠ Institutional variation: statistical learning versus computational intelligence
| UF and the statewide description | FAMU and FSU |
|---|
| Supervised and unsupervised learning | The same core |
| Linear and non-linear classification | Plus neural networks emphasised |
| Linear dimensionality reduction | Plus fuzzy logic |
| Probability and statistics for uncertain data | Plus genetic algorithms |
What this means in practice. Fuzzy logic and genetic algorithms belong to the computational intelligence tradition — a related but distinct lineage from statistical machine learning. They are genuinely useful (fuzzy control is deployed in industrial systems; evolutionary optimisation solves problems gradient methods cannot), but they are not what the term “machine learning” denotes in current industry usage, and time spent on them is time not spent on the statistical material.
A syllabus test: if membership functions and crossover operators appear in the assessed work, you are in the computational-intelligence version. Neither is wrong; a student should simply know which one their transcript line represents, particularly when applying for machine learning roles or graduate programmes that will assume the statistical reading.
⚠ The prerequisites are the real gate — and they are not programming
Requirements differ across the three institutions:
- UF: EEL 3135 (signals and systems) and EEL 3850 (probability), each with a minimum grade of C.
- FAMU/FSU: EEL 3135, MAS 3105 (linear algebra), working knowledge of MATLAB and/or Python, and instructor permission.
- Statewide: EEL 3135 or equivalent and STA 3032 or equivalent.
Read across those and the binding requirements are clear: linear algebra and probability. Machine learning is linear algebra applied to statistical estimation — least squares is a projection, PCA is an eigenvalue problem, and a neural network is a chain of matrix multiplications with non-linearities between them. Students who took linear algebra as a mechanical course (row reduction, determinants) and never worked with eigenvectors conceptually will struggle, however interested they are in the applications.
⚠ Programming fluency is assumed, not taught. FAMU/FSU state this explicitly; the others assume it. You will be expected to implement algorithms and manipulate datasets from the first assignments. A student whose programming experience is limited to short scripts should build fluency with NumPy or MATLAB matrix operations before the term, because fighting the language while also learning the mathematics is a poor use of a semester.
The instructor-permission requirement at FAMU/FSU is worth noting — it indicates a course that manages its enrolment and expects preparation to be confirmed rather than assumed. Contact the department early.
⚠ Course-code variation across Florida
- EEE 4773 — FAMU, FSU, UF. EEE 3773 — UF only (see above).
- EEE 5776 — the graduate counterpart at FAMU, FSU and UF; EEE 6778 Applied Machine Learning II at UF and USF.
- EEE 4872 Artificial Intelligence at FAMU and FSU is an adjacent but broader course covering search, natural language processing, robotics and image processing alongside learning.
- Extensive comparable material sits under CAP, CAI and COT prefixes in computer science departments, and under STA in statistics.
⚠ This is a strong prefix-divergence case. A student who took machine learning under a CAP or STA number has covered much of the same material with different emphasis — a computer science version will weight algorithms and complexity, a statistics version will weight inference and model assumptions. A receiving programme naming EEE 4773 will not match those automatically. Read target-programme prerequisites literally and carry the syllabus.
Position in the curriculum
EEE 4773 is a senior-level elective following signals and systems and probability. It pairs exceptionally well with EEE 4510 Digital Signal Processing, and students should take DSP first where the schedule allows — feature extraction from signals is where an electrical engineering graduate has an advantage over a generic machine learning candidate, and DSP is what supplies it. It leads into graduate machine learning (EEE 5776) and into research in signal processing and autonomy.
Difficulty and time commitment
The characteristic difficulty of this course is not any single topic but the combination of mathematics and implementation: a typical assignment requires deriving an estimator, implementing it, running it on real data, and interpreting a result that is worse than expected. Each of those is manageable; together they take time.
Plan on nine to eleven hours a week, more where a term project is set. Projects in this course reliably expand — data cleaning always takes longer than anticipated, and a model that underperforms invites indefinite tinkering. Set a stopping rule in advance.
Articulation and transfer
SCNS records EEE 4773 as guaranteed to transfer to an institution offering the same course, and three institutions carry it, all at 3 credits. The course is upper-division and carries no general-education or Gordon Rule designation.
FE exam relevance
The NCEES Fundamentals of Engineering (Electrical and Computer) exam covers probability and statistics as a named topic area, which this course reinforces substantially, but does not test machine learning. This is an elective taken for its own value and for employability.
AI Integration
This is a course about machine learning, so the question is not whether to discuss AI but how to use AI tools while learning the subject they implement — and the answer is more specific than in other courses.
Where a general-purpose assistant genuinely helps. Explaining a derivation such as the normal equations or the EM algorithm in different words; generating scikit-learn or MATLAB boilerplate, which is formulaic; debugging shape and broadcasting errors in NumPy, which consume an unreasonable share of a beginner’s time; suggesting visualisations; and explaining an unfamiliar method encountered in a paper.
⚠⚠ Where it fails, and why the failure is exactly this course’s subject. The characteristic error of an AI tool asked a machine learning question is to produce a working pipeline that reports an excellent accuracy figure obtained through a methodological error. This is the deepest trap in the field and it is what the course exists to inoculate against. The specific failures recur:
- Data leakage. Scaling or imputing before splitting, so test-set statistics influence training. The code runs, the accuracy rises, nothing errors.
- Test-set reuse. Tuning hyperparameters against the test set, which converts it into a validation set and destroys the only unbiased estimate you had.
- Accuracy on imbalanced data. A generated pipeline will report 97% accuracy on a dataset where 97% of examples are one class — a model that has learned nothing. Recognising this is a core learning outcome of the course, and it is precisely what a fluent generated answer omits.
- Splitting the wrong unit. Splitting by sample rather than by subject, patient or recording, so the same source appears in both training and test data. Endemic in biomedical and sensor work.
Every one of these produces a better-looking result than correct methodology would. That is what makes them dangerous: the error signal points the wrong way. A student who accepts a generated pipeline because its accuracy is high has learned the opposite of what the course teaches.
A further failure worth naming: models will recommend a method without interrogating whether its assumptions hold — suggesting naive Bayes for strongly correlated features, or linear regression for data whose residuals are plainly non-normal. Assumption-checking is the part of statistical practice that separates an engineer from a library user, and it is absent from a generated recommendation.
⚠ A note on the field’s own claims. This course is also the right place to develop scepticism about machine learning results generally, including published ones. Reproducibility problems in applied machine learning are well documented, and a large fraction trace to exactly the leakage and evaluation errors above. The habit worth forming is to ask of any reported performance: what was the split, what was the baseline, what is the class balance, and was the test set touched during development?
The engineer’s responsibility. A model deployed in a real system makes decisions about people, equipment or safety. The engineer signs for the evaluation, not for the accuracy number — and an evaluation is only meaningful if the methodology that produced it is defensible. If you cannot describe how the data were split and why the metric is the right one, you do not have a result.
Academic integrity. UF, FAMU and FSU each maintain academic integrity policies covering AI-generated work, and practice varies by instructor. Derivations, methodology decisions and result interpretation are normally expected to be your own even where coding assistance is permitted — and they are the assessed content in any case, since the code is the easy part. Generated results are data fabrication. Ask before you rely on a tool, and disclose its use where the syllabus requires it.