Machine Learning
Part of the Interview Prep Roadmap. Coverage ratings are explained there.
Optional track — relevant for ML-focused roles.
Model architecture and transformer internals, then RAG systems and training dynamics through to practical deployment considerations.
Coverage: thin
ML Fundamentals
- Supervised vs. unsupervised learning
- Training, validation, and test splits
- Overfitting and underfitting
- Bias-variance trade-off
Practice
- Supervised, unsupervised and reinforcement learning — Sonargaon University
- Overfitting and bias — IUBAT
- Overfitting vs. underfitting, drawn — Green University
- K-means clustering — Stamford University
Model Training
- Loss functions
- Gradient descent and optimizers
- Learning rate and batch size effects
- Regularization
Practice
- Batch size 8 vs. 64 — Therap
- Training on unshuffled, class-ordered data — Therap
- Scheduling models on one GPU — Enosis, with answer
Evaluation
- Accuracy, precision, recall, and F1
- Confusion matrix terms
- Metric selection for imbalanced data
- Cross-validation
Practice
- Read a training/validation curve — Therap
- Softmax on a multi-label task — Therap
- Metrics for a 99/1 imbalanced dataset — Therap
- Build a three-way news classifier — Therap
Neural Networks
- Activation functions
- Backpropagation
- Vanishing and exploding gradients
- Residual connections
Practice
Transformers & LLMs
- Attention and query/key/value matrices
- Multi-head attention
- Positional encoding
- Mixture-of-experts routing
- Temperature and sampling parameters
Practice
- Q/K/V and attention matrix shapes — Therap
- Why an argmax MoE router will not train — Therap
- Temperature when the output must be valid JSON — Therap
Applied LLM Systems
- Retrieval-augmented generation
- Chunking and embedding strategies
- Context window management
- Prompt design and few-shot examples
Practice
- Design a RAG system on a small context window — Therap
- Run inference on 120FPS CCTV footage — Therap
- Discuss a recent ML paper — Therap