← Course Home Module 11 · Oral Exam Bootcamp
Module 11 · Exam Readiness

Oral Exam Bootcamp

Everything converges here: the two papers side by side, the themes examiners cross-examine, the answering technique that earns marks, a question bank to drill aloud, and a final mock exam. Work through this module last — and more than once.

▶
Audio recap
A ~2-minute spoken summary of this module — great for revision on the go.

11.1 The two papers side by side

An oral exam on two papers almost always contains the instruction "compare them." Know this table cold — every row is a potential question.

DimensionPaper 1 — Variant classificationPaper 2 — Material recognition
ProblemClassify pathogenic SNVs / small indels across 20 monogenic diseases + "Not a Disease"Recognize materials from touch, robustly across different tactile sensors
DomainClinical genomicsRobotic tactile perception
Input101-bp DNA window, variant centered, one-hot encodedTactile images from vision-based sensors (DIGIT, GelSight, …)
ArchitectureMultiscale Conv1D → max pool → BiLSTM → attention → dense softmaxViT student (tactile encoder) aligned to a frozen BART language teacher; lightweight MLP classifier on top
Loss(es)Class-inverse-weighted categorical cross-entropyCE (student) + KL feature distillation, combined as αLstudent + (1−α)LKD with α = 0.25, T = 3.5
Scarcity / imbalance strategy500× augmentation (reverse complement, k-mer jitter) + weighted loss + stratified samplingLanguage supervision via distillation + frozen-encoder few-shot fine-tuning (0.017% of parameters trained)
EvaluationAccuracy, class-weighted F1, specificity, fallout, NPV, negative likelihood, per-class ROC/PR, confusion matrix, stratified k-foldTop-1 accuracy, macro precision/recall/F1; few-shot curves, cross-sensor transfer, six external benchmarks + ablations
ExplainabilitySHAP / PoSHAP motif maps — planned, not deliveredGrad-CAM saliency + UMAP embedding plots — delivered
Key result94.7% accuracy, F1 0.93, AUC-PR 0.98, ≈12-epoch convergence95.06% on own dataset; ~95% at 100 shots; +13.3% average cross-sensor; up to +19% across six datasets
Main limitationSynthetic random negatives & augmented backgrounds; no external validation or calibration yetVision-based tactile sensors only; language may discard fine geometric/contact detail
The 10-second contrast Paper 1 fights data scarcity by manufacturing more data (augmentation) inside one domain; Paper 2 fights it by importing knowledge from another modality (language distillation). Same enemy, opposite weapons — that sentence alone answers half the comparison questions you will get.

11.2 Shared themes examiners will probe

1 — Both are supervised multi-class classification

Softmax over classes, cross-entropy at the core, argmax at inference. Everything from Modules 1–3 (gradients, backprop, optimizers, generalization) applies identically to both. If asked "what is the mathematical task?", the answer is the same for DNA windows and tactile images.

2 — Scarce / imbalanced data: augmentation vs transfer

Compare explicitly: augmentation (Paper 1) creates label-preserving variations of what little you have — cheap, domain-knowledge-driven (reverse complement is valid because DNA is double-stranded), but risks artificial patterns. Distillation / transfer (Paper 2) borrows structure learned elsewhere (BART's semantic space) — more powerful under severe scarcity (few-shot), but requires a meaningful bridge (paired language descriptions) and inherits the teacher's blind spots. Neither replaces real data; both are bridges to it.

3 — Generalization worry: it is the SAME issue — domain shift

Paper 1's fear: trained on synthetic negatives/backgrounds → does it work on real benign variants? Paper 2's fear: trained on one sensor → does it work on another sensor? Both are distribution shift: the test distribution differs from training. Paper 2 measures its shift directly (cross-sensor benchmarks) and attacks it by design (sensor-agnostic language anchor); Paper 1 acknowledges its shift but defers measurement to future work (gnomAD negatives, prospective validation). Saying this out loud — "same underlying problem, different maturity of response" — is an A-grade synthesis answer.

4 — Interpretability

SHAP/PoSHAP (Paper 1, planned): feature-attribution — which nucleotide positions push the prediction, aggregated into positional maps a biologist can check against known motifs. Grad-CAM (Paper 2): gradient-based saliency — which image regions drove the class score. UMAP (Paper 2): not attribution at all, but a 2-D picture of embedding geometry showing tighter class clusters after distillation. Know which tool answers which question — and that UMAP distances are qualitative evidence, easy to over-read.

5 — Hyperparameters and their justification

Both papers use batch size 32, and both justify small batches by gradient noise aiding generalization (Paper 1 citing Keskar et al.; Paper 2 finding B = 32 best in its AS-4 ablation). Both use early stopping. Learning rates: 1×10−4 Adam (Paper 1) vs 2×10−5 fine-tuning (Paper 2) — small, because both fine-grained genomic features and pretrained representations are easily destroyed by large steps. Paper 2 additionally ablates α and T; Paper 1 additionally weights its loss. If an examiner asks "defend any hyperparameter," you have a rehearsed answer for either paper.

6 — Evaluation rigor

Paper 1: stratified splits + stratified k-fold cross-validation + a wide per-class metrics zoo — rigor within one (partly synthetic) dataset. Paper 2: six external public benchmarks + few-shot/zero-shot/cross-sensor regimes + modality and architecture ablations — rigor across datasets, closer to real-world evidence. Fair critique: Paper 1's internal numbers are near-perfect but internally inconsistent in places (Fig. 6, the ablation placeholders); Paper 2's evidence base is broader.

11.3 How to answer in an oral exam

Oral exams grade your reasoning process as much as your facts. Use this four-step shape for every substantive question:

  1. Direct answer first, one sentence. Examiners reward candidates who commit. "Yes — because X" beats two minutes of throat-clearing.
  2. Mechanism. Explain why, at the level of the underlying machinery (gates, gradients, distributions).
  3. Paper-specific evidence. Anchor it: an equation number, a table, a number you are sure of.
  4. Honest limits. Add the caveat before the examiner does. This converts a follow-up attack into a point you already scored.

Never bluff numbers. A confidently wrong "97.3%" is far worse than "I recall it as roughly 95%; the exact figure is in Table II." If you genuinely don't know, say so and reason aloud from first principles — examiners routinely pass candidates who derive their way to a partial answer and fail candidates who improvise facts.

Worked example — "Why a BiLSTM rather than a plain LSTM?"

Strong answer: "Because a variant's effect depends on context on both sides, and a unidirectional LSTM only ever sees one side at each step. (Direct.) An LSTM processes left-to-right, so its state at the mutation site encodes only upstream bases; a BiLSTM runs a second LSTM right-to-left and concatenates the states, so every position is represented with both its upstream and downstream context. (Mechanism.) In Paper 1 this is Equations 3 and 4 — forward and backward passes over the window — and the paper argues DNA motifs like splice signals are defined by two-sided context. (Evidence.) The cost is doubled computation and it's non-causal, which is fine here because the full 101-bp window is available at once — this isn't streaming prediction. (Limits.)"

Weak answer, for contrast: "BiLSTMs are more powerful and get better accuracy, and the paper used one because it's state of the art." — No mechanism, no evidence, an unsupported superlative, and it invites the killer follow-up "more powerful how?" which now must be answered from a position of weakness.

Three tactical habits (1) If a question is ambiguous, briefly state the interpretation you're answering: "If you mean robustness to class imbalance, then…". (2) When corrected, agree fast, integrate, and move on — defensiveness costs more than the error. (3) Silence for three seconds while you structure an answer reads as thoughtfulness, not ignorance.

11.4 The question bank — drill these aloud

No answers here on purpose: these are for self-testing out loud, ideally to another person. If a block feels shaky, the pointer tells you which module to re-read.

Warm-up definitions (→ Modules 1–2)

  1. What does softmax do, and why must its output sum to 1?
  2. Define overfitting. Name two defenses used in each paper.
  3. Why one-hot encode DNA letters instead of feeding the integers 1–4 directly?
  4. What is cross-entropy with a one-hot target, in one formula and one sentence?
  5. Precision vs recall vs specificity — one clinical sentence each.

Mechanism (→ Modules 3, 5–8)

  1. Walk through backpropagation on a two-layer network. Where does the chain rule enter?
  2. What problem do LSTM gates solve that a vanilla RNN has? Name the gates.
  3. Explain attention's Query/Key/Value in three sentences, for Paper 1's use of it.
  4. What does the temperature T do to a softmax, and why does distillation raise it above 1?
  5. How does a ViT turn an image into a sequence, and what plays the role of "words"?
  6. Why does max pooling after convolution make motif detection position-robust?

Design defense (→ Modules 9–10)

  1. Defend the 101-bp window — then attack it.
  2. Why freeze the BART teacher during distillation instead of fine-tuning it?
  3. Why class-inverse weighting in Paper 1's loss when the data was already augmented 500×?
  4. Why does Paper 2 distill at the feature level (KL between softened feature distributions) rather than matching class logits — and what did its ablation show?
  5. Both papers chose batch size 32. Give the shared justification and its honest counter-argument.
  6. Why is reverse complementation label-preserving for DNA, and what would be the analogous transformation for tactile images?

Critique (→ Modules 9–10, sections 9.7 / 10.7)

  1. What can a model trained against random-sequence negatives actually be said to have learned?
  2. Paper 1 reports ROC-AUC ≈ 1.00 for every class. Argue why that should raise, not settle, your confidence questions.
  3. What is over-reading a UMAP plot, and which of Paper 2's claims lean on one?
  4. Neither model outputs calibrated probabilities out of the box. Why does that matter more in the clinic than in robotics — or does it?
  5. Name one inconsistency in Paper 1's published figures or tables and how you would raise it diplomatically.

Synthesis (→ this module + Modules 8–10)

  1. Could Paper 1 use distillation? Sketch it: who is the teacher, what is transferred, what loss?
  2. Could Paper 2's approach work for genomics — what would the "language" be? (Think: clinical variant annotations, functional descriptions, ontology terms.)
  3. Both papers face domain shift (synthetic→real, sensor→sensor). Which paper's countermeasure would you port to the other, and how?

11.5 Final advice — the night-before checklist

The examiner's real question Behind every technical prompt sits one meta-question: "Can this person reason like a scientist about work they didn't do?" Direct answers, mechanisms, evidence, honest limits — that is the whole game.

Final Mock Exam

15 questions spanning the entire course — foundations, architectures, both papers, and cross-paper synthesis. Treat it as the real thing: no notes, first instinct, then read every explanation.

← Previous
Module 10: Paper 2 Deep Dive