D797 Artificial Intelligence and Machine Learning Foundations, catalog number ICSC 5205, is the three CU graduate course in the WGU School of Technology exploring the foundational concepts and practical applications of artificial intelligence and machine learning. Foundations is the operative word. This course is about why methods work rather than about which library call produces a model, and that distinction changes what a strong submission looks like.
Understanding the machinery you invoke
Modern tooling lets anyone train a model in a few lines, which means the ability to produce one demonstrates very little. What a foundations course assesses is whether you can explain what happened: what the algorithm is optimising, what assumptions it makes about the data, why it converges or fails to, and what a particular parameter actually controls.
The concepts that recur are the bias and variance trade off, the difference between the loss you optimise and the metric you care about, why regularisation helps, what generalisation means and why a validation set is not optional, how search and inference differ from learning, and the reason no single method is best across all problems. A submission that reasons with these rather than naming them is what the rubric is built for.
The second scored theme is honest evaluation, and it is where foundations courses catch practitioners. Any result reported without a held out comparison is unverifiable, any accuracy figure on an imbalanced target is misleading, and any model tuned against the same data used to report it has overstated itself. Getting this right is more valuable to your score than any modelling sophistication.
Turning scored aspects into a section plan
Scoring detail lives in your Course of Study rather than the catalog. Count the aspects and use them as headings, worded close to the rubric's own nouns. Each is scored independently on a three point scale and each needs a 2, so a good result with an unexplained method is a returned task.
Where a task mixes narrative with code, name the artefacts in the narrative and quote the parts that carry an aspect, since evaluators score what is visible in the document.
The word budget, worked. Assume six scored aspects and about 1,900 words of narrative. Reserve 150 words for an opening naming the problem and 120 for a close. That leaves near 1,630, about 270 per aspect. Then take 60 words out of each of two descriptive aspects and give 120 to the aspect covering conceptual explanation and 120 to the aspect covering evaluation. Explaining why a method suits the data, and evaluating honestly, are the two places a foundations submission proves it understands rather than executes.
A structure that fits a foundations task
Directions win where they specify a shape. Where they do not, this ordering matches how the aspects tend to be scored.
| Section | What belongs in it | How it gets read |
|---|---|---|
| Problem type | Whether this is classification, regression, clustering, search or inference, and why | Misclassifying the problem type invalidates everything after it |
| Data characteristics | Size, dimensionality, balance, noise and what is missing | Determines which methods are even applicable |
| Method and rationale | The approach, its assumptions, and why those assumptions hold here | The core scored reasoning in a foundations course |
| Mechanism | What the algorithm optimises and how it gets there | Explaining the machinery rather than the interface |
| Evaluation | Splits, baseline, metric chosen to match the problem | The section most likely to decide the outcome |
| Interpretation | What the result says, and what it cannot say | Overreach is the standard weakness |
| Limits and ethics | Data provenance, bias, and the cost of being wrong | Expected at graduate level even in a technical course |
Explain one parameter properly rather than listing all of them. A paragraph on what a regularisation strength actually trades away demonstrates more understanding than a table of every setting you used.
Evidence craft in a foundations course
The standard of proof here is about verifiability rather than about scale.
- Attach the data split to every number reported, without exception.
- State the baseline, including the naive one, so results can be judged.
- Report class balance for any classification problem, since it changes what a metric means.
- Cite methods to research or textbook sources rather than to library documentation.
- Record seeds and versions so the work is reproducible, and state where results vary.
- Use APA throughout and keep quotation short, since definitions are easy to lift and add nothing.
State what your data cannot support. A model trained on one population, period or context does not generalise beyond it without evidence, and saying so plainly is the cheapest credibility available in a graduate submission.
What separates Competent from a submission sent back
Aspects score independently, and returns here concentrate on explanation and evaluation.
- The method rationale refers to properties of this data rather than to general popularity.
- The mechanism is explained in terms of what is optimised and how.
- Every metric names its split and sits beside a baseline.
- Metric choice matches the problem, particularly where classes are imbalanced.
- Interpretation stays inside what the evidence supports.
Performance assessment work can be revised and resubmitted with no grade penalty, so a return costs calendar rather than standing. Terms run six months at a flat rate, and a foundations course cleared properly makes every applied course afterwards faster.
Six mistakes that cost time in D797
- Preprocessing before splitting. Fitting a transformation on all the data leaks holdout information and inflates every result.
- Accuracy on imbalanced data. It rewards predicting the majority class and hides complete failure on the class that matters.
- Tuning and reporting on the same data. The reported figure then describes the tuning rather than the model.
- Naming a method without its assumptions. The assumptions are why the method is appropriate, and they are the assessed part.
- Treating clustering output as fact. Clusters require interpretation and a stability check before they mean anything.
- Skipping the ethics discussion. Provenance, bias and cost of error are expected and take two honest paragraphs.
- Reaching for the most complex method available. A foundations course rewards the choice that fits the data, and a linear model that performs comparably to something elaborate is a finding worth reporting rather than an embarrassment to hide.
The trade off that explains most of the course
If one idea deserves extra attention in D797, it is the relationship between model flexibility and generalisation, because almost every practical decision in machine learning is an instance of it.
A model that is too simple cannot represent the pattern in the data. It performs poorly on the training set and equally poorly on new data, and no amount of additional examples will fix it, because the limitation is structural. A model that is too flexible represents the pattern and the noise together. It performs superbly on the training set and poorly on anything new, because it has memorised particulars that do not recur.
Almost every technique is a lever on this. Regularisation penalises complexity, deliberately accepting worse fit on the training data in exchange for better behaviour on new data. More training examples let you afford more flexibility, which is why the same model can be right at one data volume and wrong at another. Feature selection reduces flexibility by removing dimensions in which noise can be fitted. Early stopping halts the process before memorisation begins. Ensembling reduces the variance of an unstable model by averaging several of them.
Writing an analysis in these terms is what a foundations rubric wants. Instead of reporting that a parameter was set to a value, you can say that the model was overfitting, visible in the gap between training and validation performance, so flexibility was reduced by a stated mechanism, and the gap closed at the cost of a small drop in training accuracy. That sentence demonstrates understanding of the machinery, and it is available on almost any task in the course.
How support works on this course
Send the rubric from your Course of Study and the task directions. What comes back is aspect mapped: method rationale tied to data properties, mechanism explained in terms of what is optimised, evaluation with splits and baselines, and interpretation kept inside the evidence. Where the course includes a proctored objective assessment, support is preparation only, never sitting it, and we never ask for portal credentials.
D797 underpins the applied artificial intelligence courses in the sequence. Terms are six months at a flat rate, so a solid foundation here lowers the effort of everything downstream.
Questions students ask about D797
Is D797 the same course as ICSC 5205?
How much mathematics does D797 require?
Can you complete the modelling work for me?
Where D797 sits in WGU's programs
The July 2026 catalog places this code in 4 current WGU programs. Open a program page for the complete standard path and term positions. The live Degree Plan remains authoritative after transfer credit, substitutions, and mentor planning.
The assessments, one by one
The public catalog does not publish this course's PA/OA identity or task count. WGU Tutors publishes at most one PA manual per course and only from a WGU-controlled public rubric. Until that source exists, PA help begins from the student's real Course of Study and OA support remains preparation only.