NTU Machine Learning 2022 (Hung-yi Lee) — 15 Homeworks

Independent, from-scratch implementations of all 15 homework assignments of NTU "Machine Learning" (Spring 2022) taught by Prof. Hung-yi Lee (李宏毅), part of a csdiy.wiki full-catalog build.

Overview

This repo implements every assignment from the 2022-spring edition of the course (the canonical 15-HW deep-learning sequence: regression → classification → CNN → self-attention → Transformer → GAN → BERT → autoencoder → explainable AI → adversarial attack → domain adaptation → RL → network compression → life-long learning → meta learning).

Everything runs on CPU only (torch.set_num_threads(3), OMP_NUM_THREADS=3). Several official Kaggle datasets are competition-gated and multi-GB (food-11, VoxCeleb, Crypko anime faces, DRCD, …). Where that is the case, each HW runs the identical model / algorithm on a real, freely-downloadable dataset (CIFAR-10, FashionMNIST, MNIST, Omniglot, Multi30k), auto-downloaded via torchvision / HuggingFace datasets. Every number below is measured from a real run — see each HW's results/ for logs, metrics, and figures.

Results (measured on CPU, 3 threads)

HW	Topic	Dataset (used)	Key measured result
1	Regression (DNN)	COVID case prediction (official layout)	valid RMSE 1.46
2	Classification (BN-MLP)	phoneme frames (official layout)	frame acc 0.69
3	CNN image classification	CIFAR-10	test acc 0.6512
4	Self-attention	FashionMNIST-as-sequence	test acc 0.8540
5	Transformer (NMT)	Multi30k de→en (real)	BLEU-4 0.1560
6	GAN (DCGAN)	MNIST	pixel mean/std gap 0.006 / 0.027; sample grid
7	BERT extractive QA	multi-entity QA + real bert-tiny	EM 0.9033 / F1 0.9033
8	Autoencoder anomaly det.	CIFAR-10 (airplane=normal)	ROC-AUC 0.6203
9	Explainable AI	CIFAR-10 CNN	deletion top-k 0.307 > random 0.249 ✓
10	Adversarial attack	CIFAR-10	clean 0.539 → PGD-10 0.000
11	Domain adaptation (DaNN)	CIFAR-10 (RGB→edge domain)	target acc: src-only 0.5100 → DaNN 0.5445
12	RL (policy gradient)	CartPole (exact physics)	last-50 return 480.9 / 500
13	Network compression (KD+prune)	FashionMNIST	54.5× smaller; KD/prune curve
14	Life-long (EWC)	Permuted-MNIST	forgetting 0.027 (EWC) vs 0.224 (SGD)
15	Meta learning (MAML)	Omniglot 5-way 1-shot	MAML 0.6512 vs baseline 0.3972

Figures: hw06-gan/results/samples.png (generated digits), hw09-explainable/results/attributions.png (4 attribution methods), hw10-attack/results/adversarial_examples.png, hw12-rl/results/reward_curve.png.

Implemented assignments

• HW1 Regression — DNN regression, feature selection, official-layout data.
• HW2 Classification — BatchNorm MLP phoneme frame classifier (41 classes).
• HW3 CNN — from-scratch VGG-style CNN (no pretrained weights) on CIFAR-10.
• HW4 Self-attention — TransformerEncoder + masked mean-pool sequence classifier.
• HW5 Transformer — full encoder-decoder NMT on real Multi30k (de→en), BLEU.
• HW6 GAN — DCGAN generating real MNIST digits (non-saturating loss).
• HW7 BERT — extractive QA span prediction fine-tuning a real bert-tiny.
• HW8 Autoencoder — reconstruction-error anomaly detection, ROC-AUC.
• HW9 Explainable AI — saliency / smooth-grad / integrated-grad / occlusion + deletion metric.
• HW10 Attack — FGSM / PGD / MI-FGSM white-box L-inf attacks.
• HW11 Adaptation — Domain-Adversarial NN (gradient-reversal layer).
• HW12 RL — REINFORCE + learned baseline (actor-critic style) on CartPole.
• HW13 Compression — knowledge distillation + global L1 pruning, 54.5× smaller student.
• HW14 Life-long — Elastic Weight Consolidation on Permuted-MNIST.
• HW15 Meta learning — MAML (2nd-order) few-shot on Omniglot.

Project structure

lhy-ml-homeworks/
├── hw01-regression/ ... hw15-meta/   # one folder per homework
│   ├── hwNN_*.py                     # the implementation
│   ├── README.md                     # task + measured result
│   └── results/                      # metrics.txt, logs, figures (real runs)
├── scripts/download_data.py          # optional pre-download of all datasets
├── requirements.txt
└── LICENSE

How to run

# Python 3.11; the shared csdiy venv already has torch/torchvision/transformers/datasets:
#   D:\Project\_csdiy\.venv-ml\Scripts\python.exe
pip install -r requirements.txt

# each HW is self-contained; datasets auto-download on first run
cd hw03-cnn   && python hw3_cnn.py --epochs 8 --subset 10000
cd hw12-rl    && python hw12_rl.py --episodes 800
cd hw14-lifelong && python hw14_ewc.py --tasks 4 --epochs 3
# ... see each hwNN/README.md for its exact command

Verification

Every result was produced by an actual CPU run; each hwNN/results/ holds the run_log.txt, a metrics.txt of the measured numbers, and (where relevant) the generated figures/samples. Highlights that double as correctness checks:

• HW9 deletion metric: masking the most-salient pixels drops confidence more than masking random pixels (0.307 vs 0.249) → attributions are meaningful.
• HW10: PGD-10 drives accuracy from 0.539 to 0.000 at the exact 8/255 L-inf budget → the attack is real and correctly projected.
• HW12: last-50-episode mean return 480.9 / 500 → policy solves CartPole.
• HW14: EWC cuts catastrophic forgetting ~8× vs naive SGD (0.027 vs 0.224).

Tech stack

Python 3.11 · PyTorch 2.x (CPU) · torchvision · HuggingFace transformers & datasets · numpy · matplotlib · scikit-learn.

Key ideas / what I learned

• Building each core deep-learning primitive end-to-end: conv nets, multi-head self-attention, full Transformer encoder-decoder with masking, DCGAN training dynamics, BERT span-extraction heads.
• Beyond-accuracy topics: reconstruction anomaly scoring, gradient-based attribution + a quantitative faithfulness check, projected-gradient adversarial attacks, gradient-reversal domain adaptation, policy-gradient RL with a baseline.
• Model-lifecycle methods: knowledge distillation & pruning (compression), Elastic Weight Consolidation (continual learning), and second-order MAML (meta-learning) — each with a controlled baseline to isolate its effect.

Credits & license

Based on the assignments of NTU "Machine Learning" (Spring 2022) by Prof. Hung-yi Lee (李宏毅) — course site: https://speech.ee.ntu.edu.tw/~hylee/ml/2022-spring.php. This repository is an independent educational reimplementation; all course materials, datasets, and specifications belong to their original authors. Original code here is released under the MIT License.