Advanced Machine Learning

Assignment solutions for the Advanced Machine Learning course (IBIO-4615). This repository covers a wide range of advanced topics in machine learning, from generative networks to reinforcement learning.

---

Table of Contents

#	Topic	Folder
1	GANs	GANs/
2	Graph Neural Networks	Graphs/
3	Neural Radiance Fields (NeRF)	NeRF/
4	Recurrent Neural Networks (RNN)	RNN/
5	Reinforcement Learning	Reinforcement Learning/
6	Robustness	Robustness/
7	Self-Supervised Learning	Self Supervised/
8	Transformers	Transformers/

---

1. GANs

Implementation of Conditional GANs (CGAN) for MNIST image generation and CycleGAN for unpaired image-to-image translation (Van Gogh style to photographs). Includes ablation studies on the different CycleGAN loss functions.

CycleGAN results: Van Gogh style transfer to real photographs

---

2. Graph Neural Networks

Node classification on the OGBN-Proteins dataset using PyTorch Geometric. Several architectures are implemented and compared, including PlainGCN, GAT, ResGCN, and DenseGCN, along with hyperparameter tuning.

Deep GCN architecture: backbone, fusion, and prediction blocks

---

3. Neural Radiance Fields (NeRF)

Implementation of NeRF for novel view synthesis from 2D images, trained on the Lego dataset. Additionally explores D-NeRF for dynamic scenes using pretrained models.

---

4. Recurrent Neural Networks (RNN)

Two applications of recurrent neural networks:

• Text Generation: LSTM model trained on a jokes dataset
• Image Captioning: CNN-RNN encoder-decoder architecture on the COCO dataset

---

5. Reinforcement Learning

Implementation of classic reinforcement learning algorithms:

• Q-Learning on the Frozen Lake environment
• Deep Q-Network (DQN) on the CartPole environment

Frozen Lake environment for Q-Learning

---

6. Robustness

Study of adversarial robustness in neural networks:

• Generation of adversarial attacks (FGSM and PGD) on ImageNet
• Adversarial training of robust models on MNIST
• Comparison between standard and robust models

Adversarial attack example: an imperceptible perturbation changes the model's prediction

---

7. Self-Supervised Learning

Implementation of PiCIE (Unsupervised Semantic Segmentation using Invariance and Equivariance in Clustering) for unsupervised semantic segmentation. Explores the Siamese architecture and the effect of different transformations.

PiCIE architecture: cross-view training with invariance and equivariance

---

8. Transformers

Fine-tuning of pretrained transformers for text classification. Predicts Yelp review scores (1-5 stars) using the Hugging Face libraries, experimenting with different hyperparameters and architectures.

---

Repository Structure

Advanced-Machine-Learning/
├── GANs/                      # Conditional GANs & CycleGAN
├── Graphs/                    # Graph Neural Networks (PyTorch Geometric)
├── NeRF/                      # Neural Radiance Fields & D-NeRF
├── RNN/                       # Text Generation & Image Captioning
├── Reinforcement Learning/    # Q-Learning & DQN
├── Robustness/                # Adversarial attacks & robust training
├── Self Supervised/           # PiCIE - Unsupervised segmentation
├── Transformers/              # Transformer fine-tuning (Hugging Face)
└── Slides/                    # Course presentations