Module 0 · Foundations

Start Here: The Two Worlds

Before any math or machine learning, you need to understand the two real-world problems the papers tackle. This module assumes zero background in biology or robotics — it is written for a smart newcomer.

▶

Audio recap

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

Why this module exists An oral exam almost always opens with: "In your own words, what problem does this paper solve, and why does it matter?" If you can't answer that in plain language, nothing technical that follows will save you. This module gives you that answer for both papers.

0.1 World 1: DNA, genes, and genetic disease (Paper 1)

DNA is a very long text written in a 4-letter alphabet

Every cell in your body contains DNA — a molecule that stores the instructions for building and running you. Chemically it is a chain of four building blocks called nucleotides (or "bases"), abbreviated A, C, G, T. So you can think of DNA as a text string, about 3 billion letters long, written in a 4-letter alphabet:

...AGTAAAGGGTATGTCTATACCATAAA...

A gene is a stretch of this text (typically thousands of letters) that acts like one recipe — usually the recipe for one protein, a molecular machine that does a job in the body (carrying oxygen, digesting food, building tissue).

Mutations: typos in the text

When DNA is copied or inherited, "typos" can occur. The two kinds Paper 1 deals with:

SNV (single-nucleotide variant): one letter is swapped for another — e.g. a G becomes an A. The most common type of variant.
Indel (insertion/deletion): a few letters are inserted or deleted.

Crucially, most variants are harmless ("benign") — humans differ from each other at millions of positions. But some variants break an important recipe. A variant that causes disease is called pathogenic.

Monogenic diseases

A monogenic disease is caused by damaging variants in a single gene (mono = one). Examples used in Paper 1: cystic fibrosis (thick mucus in lungs), sickle cell anemia (misshapen red blood cells), phenylketonuria (PKU) (inability to break down one amino acid), Wilson's disease (copper buildup). There are over 7,000 such disorders, and together they affect millions of people.

Concrete example Sickle cell anemia is classically caused by a single-letter change in the hemoglobin gene: one A becomes a T, which changes one amino acid in the protein, which makes red blood cells collapse into a sickle shape. One typo, one gene, one disease — the essence of "monogenic."

The diagnostic problem

Modern DNA sequencing machines can read a patient's genome cheaply. The hard part is interpretation: sequencing a patient surfaces thousands of variants, and someone must decide which one (if any) is the pathogenic culprit. Today that is done by expert human curators cross-checking databases — slow, expensive, and successful in only roughly 25–50% of suspected genetic cases. Families often endure years-long "diagnostic odysseys."

What Paper 1 does Paper 1 builds a deep learning model that reads a short window of DNA text (101 letters, with the variant in the middle) and classifies it into one of 20 monogenic diseases — or "Not a Disease." One unified model instead of one tool per disease, aiming to make diagnosis faster, cheaper, and more consistent. The architecture combines two neural network types you'll meet in Modules 5–6: a CNN (spots short local patterns, "motifs") and a BiLSTM (reads the whole window in both directions for context).

0.2 World 2: Robots, touch, and tactile sensors (Paper 2)

Why robots need a sense of touch

A robot arm picking up objects can see with cameras, but vision alone cannot tell it how hard, soft, rough, or slippery something is, or whether a grasped object is about to slip. Humans constantly use touch for this. To manipulate objects safely — pick up an egg vs. a metal flask — robots need tactile perception: sensing through contact.

Vision-based tactile sensors

The most capable modern touch sensors (with names like GelSight, DIGIT, GelHex) work in a clever way: a soft gel pad is pressed against the object, and a tiny camera inside the sensor photographs how the gel deforms, under colored internal lighting. The output is therefore an image — a colorful picture of the contact — which is convenient, because deep learning is excellent at images. From such images a model can recognize what material it is touching (wood, steel, fabric, leather...).

The cross-sensor problem

Here's the catch: each sensor model has different optics, gel, and lighting. The same piece of wood produces very different-looking images on a DIGIT sensor vs. a GelSight sensor. A model trained on one sensor's images learns that sensor's quirks — its colors, its lighting artifacts — entangled with the actual material properties. Swap the hardware, and accuracy collapses. Every new sensor traditionally means collecting a big new dataset and retraining. This is the sensor dependence problem.

An analogy Imagine learning to read X-rays only from one specific hospital machine that tints everything green. You might accidentally learn "green smudge = fracture." At another hospital with different equipment, you fail — not because fractures changed, but because you had learned the machine, not the medicine. Robots' touch models have the same failure.

Language as the fix

Paper 2's key observation: whatever the sensor, humans describe the same touch with the same words — "rough," "soft," "glossy," "slippery." Words capture the material's essence while ignoring the hardware. Language is therefore sensor-agnostic.

What Paper 2 does Paper 2 trains a tactile image model (a Vision Transformer, Module 7) to produce internal representations that line up with the representations of a language model (BART) reading descriptions like "glossy, splotchy, hard." This teacher–student setup is called knowledge distillation (Module 8). The result: touch representations that transfer across different sensors and learn new materials from just a handful of examples (few-shot learning). They build a 39K-sample touch–vision–language dataset to do it, and beat baselines on six public benchmarks.

0.3 Why one course covers both papers

The two papers look unrelated — hospital genetics vs. robot fingertips — but underneath they share almost all of their machinery, which is exactly what your exam will probe:

Shared concept	Paper 1 (DNA)	Paper 2 (Touch)
Input data	DNA text window (a sequence)	Tactile image
Core task	Multi-class classification (20 diseases + control)	Multi-class classification (materials)
Neural architecture	CNN + BiLSTM	Vision Transformer + BART
Loss function	Cross-entropy (class-weighted)	Cross-entropy + KL distillation loss
Data scarcity trick	Data augmentation (500×)	Language supervision + few-shot transfer
Generalization worry	Rare disease classes, synthetic negatives	Unseen sensors, unseen materials
Explainability	SHAP / motif maps	Grad-CAM / UMAP

Master the shared machinery once (Modules 1–8), then each paper becomes a case study (Modules 9–10) rather than a mountain.

Exam warm-up — say it out loud Before taking the quiz, answer these aloud in under a minute each, with no notes:

What is a pathogenic SNV, and why is finding one hard in practice?
Why does a tactile model trained on one sensor fail on another?
In one sentence each: what is Paper 1's solution? What is Paper 2's?