D802 Deep Learning, catalog number ICSC 6204, is the three-CU graduate course in the WGU School of Technology that goes underneath the libraries: the principles, the mathematics and the implementation detail of neural networks. The distinguishing demand is that you can say why an architecture, a loss function and a training schedule fit the problem, and can show from your own training curves what happened rather than reporting a final number. A network that trains is not the deliverable. An explanation an evaluator can follow, supported by evidence you generated, is.
What ICSC 6204 is actually testing
Architecture justification comes first. Convolutional structure exploits spatial locality and weight sharing; recurrent and attention-based structures exist because order and long-range dependence matter in sequences; a plain fully connected network is the right answer more often than students expect on small tabular problems. Graduate work states which property of the data motivated the structure chosen. Copying an architecture from a tutorial and describing what it contains, layer by layer, answers a question nobody asked.
Training behaviour comes second, and this is what separates deep learning from the earlier machine learning course. Loss curves are evidence. A training loss that falls while validation loss climbs is overfitting with a timestamp on it. A loss that plateaus immediately points at learning rate, initialisation or a target that was never learnable from those features. A validation curve that jumps around suggests a batch size or a learning rate fighting the data. Reporting what you saw during training, and what you changed in response, is the part of the write-up that demonstrates you understand the machinery.
The mathematics is tested through use rather than through recitation. Backpropagation as the chain rule applied across layers, the reason a saturating activation stalls gradients, what a regulariser penalises, what a normalisation layer stabilises: these appear in strong submissions as one-line explanations attached to decisions, not as a separate theory section.
Turning scored aspects into a write-up plan
The scoring detail for D802 sits in your Course of Study. Read it before you write a line of code, because the aspects usually determine what you need to have recorded during training, and training runs are hard to reconstruct after the fact. Each aspect is scored alone, and a 2 in each is what passes the task.
Keep a lab notebook from the first run: date, configuration, dataset version, what changed, what happened. Students who improvise this at the end lose a day rebuilding history, and the experiment-tracking aspect is one where reconstructed detail tends to be visibly thin.
The word budget, worked. Take six scored aspects and directions asking for roughly 2,200 words. Reserve 140 for an opening naming the task and the data and 110 for a closing on what you would try next. That leaves about 1,950, or 325 words an aspect. Now shift weight toward the two aspects that carry the reasoning, usually architecture rationale and training analysis: take 55 from each of the three descriptive aspects and give 80 extra to each of those two. Figures help here more than in most courses. A single loss curve with a two-sentence reading of it can satisfy more of an aspect than four hundred words of narration.
Where compute limits shaped your work, say so inside the relevant aspect rather than apologising in a footnote. Choosing a smaller model because you had one consumer graphics card is a legitimate engineering decision when it is stated as one.
A structure that fits a deep learning report
Task directions come first where they prescribe a layout. Otherwise this arrangement matches how deep learning aspects are usually scored.
| Section | What belongs in it | What earns the aspect |
|---|---|---|
| Task and data | What is predicted from what, dataset size, splits and preprocessing | Scored for enough detail to judge whether the results are plausible |
| Architecture | Layers, sizes, activations and the data property motivating each choice | Scored for rationale; a layer list without reasons is a description |
| Training setup | Loss, optimiser, learning rate schedule, batch size, epochs, regularisation, seed | Scored for reproducibility; missing hyperparameters make results unverifiable |
| Training behaviour | Curves, what they showed and what you changed in response | Scored for diagnosis; this is where understanding is visible |
| Evaluation | Held-out performance with metrics chosen for the task, plus a baseline | Scored for validity; test data touched once, at the end |
| Error analysis | Which examples fail, and any pattern in them | Scored for insight; aggregate metrics alone leave this aspect thin |
| Limitations and ethics | Data provenance, bias risk, compute cost, deployment constraints | Scored where named; specifics beat general concern |
| References | Papers behind the architecture and techniques, dataset citation, APA | Scored where citation is named; borrowed architectures need attribution |
If you adapted a published architecture or a tutorial implementation, say which one and what you changed. Attribution costs a sentence, and its absence in a graduate course is a serious problem rather than an oversight.
Evidence craft when the evidence is a training run
Your evidence in this course is mostly generated, which puts the burden of honesty on you rather than on a citation.
- Report the run you submitted, not the best run you remember. If several configurations were tried, show the comparison in a table.
- Label every curve: axes, units, which split, which configuration. An unlabelled loss plot cannot support a claim.
- Give hyperparameters in full, in one place. Learning rate, batch size, epochs, optimiser settings, regularisation strength, seed, hardware.
- Hold the test set back until the end and say so. A model selected on test performance has no honest test score left.
- Show at least a handful of concrete failures. In vision or language work, individual misclassified examples reveal more than another decimal place of accuracy.
- Cite architectures, techniques and datasets in APA, and state the licence where a dataset carries one.
A result you did not expect is worth more than a tidy one when you explain it. A model that beats your baseline by a suspiciously wide margin deserves a paragraph checking for leakage before it deserves a celebration, and writing that paragraph is exactly the behaviour a graduate rubric is built to reward.
What separates Competent from a submission sent back
Deep learning returns cluster in two places: an architecture described but not justified, and results reported without the training story that makes them credible.
- Every architectural choice ties to a property of the data or the task.
- Hyperparameters are complete enough for someone to rerun the experiment.
- Training curves are shown and read, not merely included.
- Overfitting is addressed with evidence, whether it appeared or was prevented.
- Evaluation uses a held-out set and a baseline, with metrics suited to the task.
- Borrowed code and architectures are attributed precisely.
Work assessed by performance assessment can be revised and resubmitted with no grade penalty, so the real cost of a return is the calendar. In a six-month flat-rate term, a three-CU course that drifts across two return cycles takes the space a fourth course could have used. Where the course includes a proctored objective assessment, we prepare only, and we never sit assessments or ask for portal credentials.
Six mistakes that cost time in D802
- Reaching for the biggest model available. On a few thousand examples, a deep architecture memorises. Capacity is a decision to justify, not a default.
- Training without a validation split. Without it there is no way to detect overfitting and no honest basis for stopping.
- Reporting only final accuracy. The curve is where the reasoning lives; the endpoint alone tells an evaluator almost nothing.
- Leaving augmentation or normalisation undescribed. These change results substantially and their absence from the write-up makes the numbers unreproducible.
- Explaining backpropagation in general terms. The rubric wants the concept used on your model, not a textbook paragraph.
- Starting the training runs in the last week. Runs fail, data loaders break and a single bad configuration can burn a day. Deep learning courses punish compressed schedules harder than most.
How support works on this course
Send the task directions, the scoring detail from your Course of Study and whatever you have run so far, including the curves. What comes back is an architecture rationale written in the terms the rubric asks for, a hyperparameter table that makes your work reproducible, a reading of your training behaviour that explains what the curves show, and an error analysis plan that turns a metric into an argument.
Where the blockage is conceptual rather than editorial, we work through the mechanism directly, whether that is why gradients vanish through a particular activation or why a normalisation layer changed your training stability, so the explanation in your document is one you can defend.
Questions students ask about D802
Is D802 the same course as ICSC 6204?
Should I take D801 before D802?
Do I need a powerful computer for this course?
Network trained and the write-up will not explain itself?
Send your task directions, your rubric and your curves. You get an architecture rationale, a complete hyperparameter table and a training analysis an evaluator can follow.
Where D802 sits in WGU's programs
The July 2026 catalog places this code in 1 current WGU program. 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.