A ground-up course preparing you for the PhD comprehensive oral exam on two research papers: Paper 1 — a windowed CNN–BiLSTM framework for pathogenic genetic variant classification, and Paper 2 — language-guided representation learning for cross-sensor tactile material recognition.
We start from first principles — no prior ML background assumed — and climb, module by module, until you can defend every design choice in both papers.
How to use this course: go in order, starting with Module 0 — it assumes no biology, robotics, or ML background at all. Each module is short, opens with a ~2-minute audio recap you can listen to for quick revision, and ends with a quiz — score at least 80% before moving on, because every later module assumes the earlier ones. Quiz scores are saved on this device and shown on each card.
No biology or robotics background needed. Plain-language primers on DNA, genetic disease and diagnosis (Paper 1's world) and on robots, touch sensors and why machines need to feel (Paper 2's world) — and what each paper set out to solve.
Vectors, matrices, functions, derivatives & gradients, probability, logs and softmax — the entire mathematical toolkit the papers rely on.
What "learning from data" means: features, labels, classification, loss functions, gradient descent, train/validation/test splits, overfitting.
Neurons, activations, layers, forward pass, backpropagation, cross-entropy loss, the Adam optimizer, regularization and early stopping.
DNA as data, one-hot encoding, embeddings, tactile images, class imbalance, and data augmentation — the raw material both papers work with.
Filters, convolution, Conv1D as a motif detector on DNA, pooling, 2D CNNs for images, and Grad-CAM for seeing what a CNN looks at.
Why sequences need memory, vanishing gradients, the LSTM gates, bidirectionality, and why CNN + BiLSTM is a natural pairing.
Attention, Vision Transformers (ViT), language models (BART), embedding spaces, and visualizing representations with UMAP.
Teacher–student training, softmax temperature, KL divergence, combined losses, frozen encoders, fine-tuning, and few-shot learning.
The full pipeline: 101-bp windows, 500× augmentation, weighted loss, stratified k-fold, ROC/PR/specificity results, limitations and explainability.
Cross-sensor tactile perception, language as a sensor-agnostic teacher, the 39K-sample dataset, few-shot & cross-sensor results, ablations.