Week 10 · PBL presentation week

PBL1 Presentations: Neural Prediction Projects

How can a group make a clear, bounded claim from its data and model evidence?

Both 100-minute classes are reserved for the first project presentations, as in the taught course schedule. Groups present a feasible prediction question, data provenance and exploration, a baseline, a compact neural comparison, held-out evidence, failures, limitations, and contributions. The chapter supports presenters and listeners; it does not add a new technical unit.

Both 100-minute classes are reserved for presentations, questions, peer learning, and reflection. No additional technical chapter is assigned this week.

By the end

Present one coherent question-to-evidence argument.
Explain why the neural model is or is not justified over the baseline.
Answer questions about splits, metrics, failures, and limitations.
Use peer projects to identify one improvement for the next project.

Low floor

Use one shared evidence spine so presenters and listeners know what to look for.

High ceiling

Defend one modeling decision by comparing it with a credible alternative and identifying evidence that could reverse the decision.

Presentation contract

Three commitments

Claim–evidence alignment

A claim must be no broader than the held-out evidence and population represented by the data.

Comparative modeling

A neural result becomes informative when compared with a transparent baseline under the same split, preprocessing, budget, and metric.

Oral accountability

Every member should be able to explain the complete workflow, including decisions outside their assigned implementation task.

Two consecutive classes

Use the full presentation week

First class · 100 minutes

Opening evidence reminder and order check.
First half of group presentations.
Questions distributed across group members.
Listeners record one strength, one limitation, and one constructive question per project.

Second class · 100 minutes

Brief restart and unresolved question.
Remaining group presentations.
Cross-project comparison of baselines, splits, metrics, and failures.
Individual reflection and next-step record.

Presenter path

Five parts of a defensible presentation

1. Open with the question, observation unit, target, intended use, and non-use.
2. Show provenance, EDA, and the frozen data split before model results.
3. Compare the baseline and neural model under aligned conditions.
4. Show at least one representative failure and one limitation.
5. Close with the bounded conclusion and answer questions using visible evidence.

Questions that help

Listen for evidence gaps

Metric parade

What goes wrong: Many metrics appear without a decision or comparison.

Ask or check: Ask which one metric answers the stated question and why.

Test data used during development

What goes wrong: The final estimate is optimistic.

Ask or check: Ask when the test set was first opened and what was frozen beforehand.

Unequal participation

What goes wrong: Only one member can explain the model or evidence.

Ask or check: Invite questions across members and use the contribution record privately.

Listener record

Learn across projects

For each project

What was the question?
What was the strongest evidence?
What important limitation remained?
What one question would improve the claim?

After both classes

Which baseline was most informative?
Which failure analysis changed your view?
Which design choice would you reuse?
What will you change in your next project?

Check your understanding

Ten questions with standard answers

Answer in your own words before opening each panel.

1. What is the first claim a presentation should establish?

Standard answer: State the prediction question, observation unit, target, intended use, and non-use before discussing architecture.

2. Why show the baseline before the neural model?

Standard answer: The baseline establishes the minimum useful comparison and shows whether added complexity changes the evidence.

3. What makes a train/validation/test split credible?

Standard answer: Its roles are fixed in advance, relevant identity or time boundaries are respected, and the test set remains closed during model selection.

4. What should one main result slide contain?

Standard answer: One bounded claim, a clearly labeled held-out comparison, and the conditions under which the comparison was made.

5. Why include a failure case?

Standard answer: It reveals where aggregate metrics hide systematic weakness and helps bound intended use.

6. What should every group member understand?

Standard answer: The full question-to-evidence path, even when implementation tasks were divided.

7. How should listeners take notes?

Standard answer: Record the question, strongest evidence, one limitation, and one constructive question rather than transcribing slides.

8. What is a useful oral-defense answer?

Standard answer: A direct answer tied to the shown data or computation, followed by what remains uncertain.

9. When is a neural model not justified?

Standard answer: When it does not improve decision-relevant held-out evidence enough to offset added complexity, instability, or opacity.

10. What should carry forward to PBL2?

Standard answer: A clearer question, frozen evidence contract, stronger baseline, deliberate failure test, and concise contribution record.

Terminology

Glossary

Presentation claim: Main conclusion defended by the project evidence.
Evidence chain: Connected path from question and data to evaluation and limitation.
Baseline: Simple comparison model.
Held-out evidence: Results on data not used to fit or select the model.
Error case: Specific example where the model fails.
Uncertainty: What is not known precisely from the available evidence.
Limitation: Boundary on the conclusion.
Contribution: Work and understanding attributable to a member.
Peer review: Structured feedback from another learner.
Oral defense: Questions testing understanding and evidence quality.

Go further

Key references

Next: Week 11 resumes technical study with a compressed comparison of RNN and LSTM sequence models.