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.
An oral exam on two papers almost always contains the instruction "compare them." Know this table cold — every row is a potential question.
| Dimension | Paper 1 — Variant classification | Paper 2 — Material recognition |
|---|---|---|
| Problem | Classify pathogenic SNVs / small indels across 20 monogenic diseases + "Not a Disease" | Recognize materials from touch, robustly across different tactile sensors |
| Domain | Clinical genomics | Robotic tactile perception |
| Input | 101-bp DNA window, variant centered, one-hot encoded | Tactile images from vision-based sensors (DIGIT, GelSight, …) |
| Architecture | Multiscale Conv1D → max pool → BiLSTM → attention → dense softmax | ViT student (tactile encoder) aligned to a frozen BART language teacher; lightweight MLP classifier on top |
| Loss(es) | Class-inverse-weighted categorical cross-entropy | CE (student) + KL feature distillation, combined as αLstudent + (1−α)LKD with α = 0.25, T = 3.5 |
| Scarcity / imbalance strategy | 500× augmentation (reverse complement, k-mer jitter) + weighted loss + stratified sampling | Language supervision via distillation + frozen-encoder few-shot fine-tuning (0.017% of parameters trained) |
| Evaluation | Accuracy, class-weighted F1, specificity, fallout, NPV, negative likelihood, per-class ROC/PR, confusion matrix, stratified k-fold | Top-1 accuracy, macro precision/recall/F1; few-shot curves, cross-sensor transfer, six external benchmarks + ablations |
| Explainability | SHAP / PoSHAP motif maps — planned, not delivered | Grad-CAM saliency + UMAP embedding plots — delivered |
| Key result | 94.7% accuracy, F1 0.93, AUC-PR 0.98, ≈12-epoch convergence | 95.06% on own dataset; ~95% at 100 shots; +13.3% average cross-sensor; up to +19% across six datasets |
| Main limitation | Synthetic random negatives & augmented backgrounds; no external validation or calibration yet | Vision-based tactile sensors only; language may discard fine geometric/contact detail |
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.
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.
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.
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.
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.
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.
Oral exams grade your reasoning process as much as your facts. Use this four-step shape for every substantive question:
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.
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.
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.