Machine Learning Experiments for Researchers

Practical companion to the IMT Machine Learning course.

This course teaches how to design, run, and audit machine learning experiments in a research-grade setting. We move from a minimal training loop to a structured, reproducible experiment pipeline.

The focus is on:

• reproducibility,
• controlled comparisons,
• structured logging,
• and experimental rigor.

---

Syllabus

1. 🌍 Big Picture
   Why ML experiments are hard today: scale, brittleness, infrastructure

2. 💻 Dev Setup in 2026
   A minimal researcher stack: IDE + AI Assist, Git, environments, tracking

3. 🔁 Training Script (Vanilla)
   Build the minimal loop: data → model → loss → optimizer → eval

4. 📊 Training Script (Research-Grade)
   Make runs comparable: configs, logging, checkpoints, run grids, basic HPO

5. 🧾 (Optional) Working with Text
   Run a tiny Transformer experiment (tokenization, batching, evaluation)

6. ⚡ (Optional) Hardware for ML
   Scope experiments: VRAM/RAM/disk, throughput bottlenecks, GPU selection

---

Course Materials

Slides

mle4r-winter26.pdf

Notebooks

#	Section	Notebook
1	Data	1_data.ipynb
2	Model	2_model.ipynb
3	Optimizer + Loss	3_optimizer_and_loss.ipynb
4	Training Loop	4_training_loop.ipynb
5	Training Script	5_training_script.ipynb
6	Transformers	6_transformers.ipynb

Scripts

• train_mnist.py - single reproducible run
• runner_simple.py - programmatic run launcher
• runner_full.py - small run grid scheduler
• hp_opt.py - minimal random hyperparameter search

These illustrate the progression:

One run
↓
Configurable script
↓
Run grid (seed × hyperparameters)
↓
Structured hyperparameter search

All materials are self-contained and runnable locally (CPU or single GPU).

---

Location and Timetable

📍 Location
IMT School for Advanced Studies Lucca
San Francesco Complex
Classroom 2

🗓 Timetable

Day	Date	Time
Friday	February 13, 2026	09:00–11:00
Monday	February 16, 2026	09:00–11:00