Applied Deep Learning
ECBS5200 · Central European University · Spring 2026
A hands-on graduate course in post-training deep learning engineering. Fine-tune, adapt, analyze, compress, and justify a real model under real constraints.
Over six weeks, you work on one cumulative problem: classifying consumer financial complaints into 113 categories. You fine-tune a pretrained encoder, improve it, adapt it with LoRA, analyze where it fails, compress it via quantization, and distill knowledge from a stronger (but 1000x more expensive) decoder reference system. At the end, you write a recommendation: is the cheap model good enough, or is the expensive one worth the cost?
Instructor: Eduardo Ariño de la Rubia · 6 Wednesdays · Apr 8 – May 13, 2026 Compute: Free-tier GPU notebooks (Kaggle T4) · Base model: ModernBERT-base (149M params)
Course Schedule
| Week | Topic | Date | Materials |
|---|---|---|---|
| — | Pre-work | Before Apr 8 | 8 modules |
| 1 | Fine-tuning and data audit | Apr 8 | Released |
| 2 | Controlled improvement and failure analysis | Apr 15 | Released |
| 3 | Parameter-efficient adaptation (LoRA/PEFT) | Apr 22 | Released |
| 4 | Error diagnosis: slices, calibration, and cross-model analysis | Apr 29 | Released |
| 5 | Quantization and decoder economics | May 6 | Released |
| 6 | Distillation and final model decision | May 13 | Released |
Latest from the Blog
Distillation acted like a calibration regularizer. There was a cheaper one. — I’ve been excited about Week 6 of this course — distillation — since the start of the term. Six weeks of fine-tuning, comparing, diagnosing, and compressing, and the experiments kept turning up th