Week 14 · PBL presentation week
What does the final project demonstrate, and where must its claims stop?
Both 100-minute classes are reserved for final project presentations and synthesis. Projects may use convolutional, recurrent, attention, Transformer, or transfer-learning ideas, but the standard is not architectural novelty. Each group must connect its question, data, baseline, model choice, held-out evidence, failures, reproducibility, and limitations into one defensible argument.
Both 100-minute classes are reserved for presentations, questions, peer learning, and reflection. No additional technical chapter is assigned this week.
Use one shared evidence spine so presenters and listeners know what to look for.
Defend one modeling decision by comparing it with a credible alternative and identifying evidence that could reverse the decision.
Presentation contract
An advanced architecture should be chosen because its inductive structure matches images, sequences, attention relationships, or transfer needs—not because it is fashionable.
Another informed reader should be able to reconstruct the split, preprocessing, model, settings, metrics, and limitations from the submitted record.
The conclusion stops at the represented population, observed task, held-out evidence, and tested failure conditions.
Two consecutive classes
Presenter path
Questions that help
What goes wrong: The group treats using a Transformer or pretrained model as the main result.
Ask or check: Ask what held-out comparison justifies the added method.
What goes wrong: Pretraining or external data are assumed to transfer automatically.
Ask or check: Document source objective, target population, overlap risks, and target-domain errors.
What goes wrong: A small course dataset is used to support a population-wide or causal claim.
Ask or check: Rewrite the conclusion to match the actual sample and evaluation design.
Listener record
Check your understanding
Answer in your own words before opening each panel.
Standard answer: Held-out decision-relevant evidence, stability, resource cost, failure behavior, and fit to the data structure—not novelty alone.
Standard answer: It shows whether the advanced method contributes enough to justify its complexity.
Standard answer: Source model and objective, data provenance where known, frozen or updated components, target adaptation, and mismatch or contamination risks.
Standard answer: A subset tied to the intended use or known risk, large enough to interpret cautiously, and defined before viewing final test results when possible.
Standard answer: Recorded data version, split, preprocessing, seed, code, model configuration, training settings, metric definition, and environment.
Standard answer: Show variation across seeds or runs, avoid selecting only the best result, and reduce the strength of the claim.
Standard answer: Prediction estimates an outcome from patterns; it does not establish that changing an input will cause the outcome to change.
Standard answer: Whether each member understands the choices, computations, evidence, failures, and boundaries of the complete project.
Standard answer: When data coverage, evaluation validity, uncertainty, or failure evidence is insufficient for the intended decision.
Standard answer: Define the problem and data boundary, establish a baseline, build the smallest justified model, evaluate held-out behavior, inspect failures, document provenance, and state a bounded conclusion.
Terminology
Go further
Next: The semester closes with a reusable discipline: define the data contract, establish a baseline, train visibly, evaluate held-out behavior, inspect failures, and state a bounded conclusion.