| Responsible AI and MLOps Foundations |
One deck in seven parts. Lecture 1: bias and fairness, interpretability (SHAP, LIME, Grad-CAM), documentation and governance (SDAIA, PDPL). Lecture 2: from experiment to production, lifecycle and versioning, deployment and monitoring, compression and efficient inference |
Open |
| Lab 1: Quantisation, Fewer Bits per Number |
Exercise: quantise Day 2's Fashion-MNIST CNN by hand (symmetric vs affine, per tensor vs per channel, 8 bits down to 2), run it in int8 with ONNX Runtime, then quantise YOLO11 and find out why quantising everything breaks it |
Notebook Colab |
| Lab 2: Pruning and Distillation, Then Ship One |
Exercise: prune the CNN to 95% sparsity and recover it, distil it into a 4x smaller student from 1,000 labels, then pick the release through a gate on per-class accuracy, size and latency |
Notebook Colab |
| Live Demo: Compressed YOLO on Your Webcam |
On your own laptop, not Colab: YOLO11m at float32 and int8, the smaller YOLO11n, and float16 on a GPU, switched live with keys 1 to 4. Run python yolo_live_demo.py |
Open |
| Optional: Quantisation (Solution) |
Worked solution to Lab 1 |
Notebook Colab |
| Optional: Pruning and Distillation (Solution) |
Worked solution to Lab 2 |
Notebook Colab |
| Optional: Audit, Explain and Document a Model |
Exercise: audit a credit pre-screening model on UCI Adult: per-group metrics with error bars, the fairness conflict, a proxy test, threshold mitigation, SHAP and LIME, and a generated model card |
Notebook Colab |
| Optional: Audit, Explain and Document a Model (Solution) |
Worked solution to the audit exercise |
Notebook Colab |
| Optional: Deploy, Monitor and Retire a Model |
Train, track, tune, register, deploy, monitor, roll back and retire a model with MLflow, Optuna and Gradio, entirely on your own machine: no accounts |
Notebook Colab |
| Optional: Responsible AI and Model Interpretability (Full Deck) |
Reading: the full deck that Lecture 1 draws from, with all four opening cases, every source of bias and the full fairness and governance sections |
Open |
| Optional: MLOps Foundations (Full Deck) |
Reading: the full deck that Lecture 2 draws from, including hyperparameter optimisation with Optuna, containers, and CI/CD for ML |
Open |
