Robust ML Pipelines (IML26 Assignment 2)

Three ML tasks focused on generalization and robustness, not overfitting.

Final score: 98.7 / 100

---

Results

• Task 1 (Image Classification): 100 / 100
• Task 2 (Tabular Classification): 100 / 100
• Task 3 (Anomaly Detection): 96.2 / 100

---

Approach

• Task 1: removed spurious corner signal → model learns real patterns
• Task 2: feature selection (top-k) + gradient boosting
• Task 3: compact NN (≤500 params) + imbalance-aware training

---

Stack

• PyTorch
• scikit-learn
• NumPy / pandas

---

Structure

• code/
• ├── train_task1.py
• ├── train_task2.py
• ├── train_task3.py

Key Idea

Generalization > fitting the train set.