AI Masters Projects

A collection of academic projects completed during the AI Masters program at Umeå University.

---

📁 Projects Overview

Project	Description	Technologies
Answer Set Programming	Logic programming for argumentation frameworks	ASP, Clingo, Python
Data Privacy 1	Masking methods (generalization, rank swapping, noise) and privacy-utility tradeoff on Adult Census	Python, scikit-learn, Jupyter
Data Privacy 2	Seminar on privacy risks in model explanations (Shokri 2021) and reconstruction attacks on unlearning (Bertran 2024)	Python, PowerPoint, Research
Data Privacy 3	Differential privacy and CTGAN tabular synthetic data generation with utility preservation benchmarks	Python, PyTorch, SDV, CTGAN
Deep Learning 1	CNN-based facial emotion recognition on AffectNet with training and evaluation scripts	Python, TensorFlow, Keras
Deep Learning 2	Manual backpropagation and gradient descent in JAX for regression and classification models	Python, JAX, matplotlib
Deep Learning 3	Conditional VAE for facial expression generation and latent-space interpolation	Python, TensorFlow, Keras
Fashion Article Classification	Image classification using k-NN and neural networks	Python, scikit-learn, TensorFlow
LLM Backend	FastAPI backend for educational game with LLM-powered hints	Python, FastAPI, Hugging Face
Machine Learning 1	Supervised learning assignment with classification and regression	Python, scikit-learn, Jupyter
Machine Learning 2	Neural networks, PCA, and clustering across California Housing, Wine, and MNIST	Python, scikit-learn, TensorFlow, Jupyter
Machine Learning 3	Human activity recognition with classical models and raw-signal deep learning	Python, scikit-learn, TensorFlow, Jupyter
MLOps-Pipeline	End-to-end MLOps pipeline for customer churn prediction with CI/CD and monitoring	Python, XGBoost, MLflow, Docker, FastAPI
Monocular Depth Estimation	YOLO26 + Depth Anything 3 fusion for real-time obstacle avoidance on an ESP32-C3 robot	Python, PyTorch, FastAPI, BLE
Natural Language Processing 1	spaCy pipeline exploration: tokenization, POS tagging, dependency parsing, NER, and multilingual Swedish NLP	Python, spaCy
Natural Language Processing 2	20 Newsgroups text classification with NLTK preprocessing, TF-IDF, and supervised linear/NB models	Python, scikit-learn, NLTK
Natural Language Processing 3	Probing vs fine-tuning Pythia language models (160M–2.8B) for implicit state changes in procedural text (bAbI & ProPara)	Python, PyTorch, Transformers, Pythia
Othello	Game AI engine with Alpha-Beta pruning	Python, Minimax, IDS
Reinforcement Learning	Bandit benchmarking and multi-agent Pong with Q-learning	Python, pytest, ma-gym
Spin the Wheel	Full-stack web application prototype	Spring Boot, Angular, Firebase
Statistics	Statistical inference analysis	R, tidyverse

---

📂 Project Details

🧠 Answer Set Programming

Course: AI-2 (5DV181) – Logic Programming

Implementation of argumentation semantics using Answer Set Programming (ASP). Includes conflict-free sets, admissible sets, stable extensions, and preferred extensions. Features both declarative ASP encodings and an imperative Python solver for performance comparison.

🔒 Data Privacy 1

Course: Data Privacy (5DV241)

Empirical evaluation of data protection and anonymization methods applied to the UCI Adult Census income dataset (~32,000 records). Investigates the tradeoff between disclosure risk and downstream utility:

• Data cleaning, quasi-identifier identification, and preprocessing
• Masking techniques: Generalization (interval rounding), rank swapping, and additive continuous noise
• Disclosure risk quantification using distance-based record linkage
• Information loss assessment: Generic statistical preservation vs specific task utility (Random Forest classification predicting income > $50K)
• Comparative tradeoff curves across four parameter configurations per method

🔍 Data Privacy 2

Course: Data Privacy (5DV241)

Advanced seminar and critical analysis of privacy vulnerabilities in machine learning models and post-hoc explanations:

• Seminar presentation and report on Shokri et al. (IEEE S&P 2021), "On the Privacy Risks of Model Explanations", investigating membership inference attacks exploiting feature explanations (SHAP, saliency maps)
• Evaluation of reconstruction attacks on machine unlearning based on Bertran et al. (NeurIPS 2024), auditing weight differences before and after deletion requests
• Technical report detailing attack models, privacy-preserving explanation mechanisms, and formal mitigation strategies
• Slide deck and structured discussion Q&A

🛡️ Data Privacy 3

Course: Data Privacy (5DV241)

Empirical investigation of Differential Privacy (DP) and generative synthetic data generation using Conditional GANs:

• Differential Privacy mechanisms across varying privacy budgets ($\epsilon$)
• Synthetic tabular data generation with CTGAN (SDV) across sample sizes ($N \in [10, 16512]$) and multiple random seeds
• Downstream utility evaluation on Logistic Regression, Random Forest, and MLP models
• Evaluation of hybrid DP-synthetic pipelines assessing compound accuracy degradation
• Scaling benchmarks and publication-ready visualization suite

😃 Deep Learning 1

Course: Deep Learning

Facial emotion recognition using a Convolutional Neural Network (CNN) trained on the AffectNet dataset. Includes:

• Custom CNN architecture implemented in Keras
• Four-class emotion recognition: Happy, Sad, Surprised, and Mad
• Training, validation, and test evaluation on AffectNet emotion data
• Visualization of loss and accuracy curves
• Provided environment YAML for reproducibility

Model file: affectnet_emotion_cnn.keras

🧮 Deep Learning 2

Course: Deep Learning (5DV236)

Backpropagation assignment implemented with JAX and the course jaxon library. The project focuses on composing layer Jacobians into full gradients and then training models with manual gradient descent.

• Manual backpropagation through model layers
• Gradient descent for a linear regression example
• Gradient descent for a feed-forward multiclass classifier
• Learning-curve generation and written report

🎭 Deep Learning 3

Course: Deep Learning (5DV236)

Conditional Variational Autoencoder (CVAE) for facial expression generation on AffectNet. The model learns a 128-dimensional latent space conditioned on four emotion classes (Happy, Sad, Surprised, Mad) and supports:

• Reconstruction of 112×112 face images with multi-level label injection in the decoder
• Expression transfer — encoding a Happy face and decoding with Sad conditioning
• Smooth latent interpolation between mean class embeddings
• β-VAE training (β = 0.15) with KL annealing to prevent posterior collapse

👕 Fashion Article Classification

Course: Machine Learning

Classification of fashion items from the Fashion-MNIST dataset using:

• Custom k-Nearest Neighbors (k-NN) implementation
• Dense Neural Network (Multilayer Perceptron)
• Hyperparameter tuning and model evaluation

🤖 LLM Backend

Standalone Project

Backend API for an educational game platform, built with FastAPI. Provides endpoints for curriculum, game logic, user modeling, and AI-powered hint generation using large language models (LLMs) via Hugging Face.

• RESTful API for game and curriculum management
• Integration with Llama-3-8B-instruct via Hugging Face Inference API
• In-memory session storage for local development
• Docker and local run support

📘 Machine Learning 1

Course: Machine Learning (5DV238)

Foundational supervised learning assignment covering:

• Classification on the Breast Cancer Wisconsin (Diagnostic) dataset
• Regression on the California Housing dataset
• Model training, evaluation, and comparison using scikit-learn

📗 Machine Learning 2

Course: Machine Learning

Second machine learning assignment covering neural network regression on California Housing, PCA on Wine and MNIST, and custom/scikit-learn K-Means clustering analysis.

📙 Machine Learning 3

Course: Machine Learning (5DV238)

Third machine learning assignment focused on human activity recognition using the UCI HAR smartphone dataset. The project compares classical models on engineered features with raw inertial-signal experiments, including a 1D CNN for sequence modeling.

⚙️ MLOps-Pipeline

Standalone Project

End-to-end MLOps pipeline for customer churn prediction using XGBoost. Demonstrates production-grade ML engineering practices:

• Experiment tracking and model registry with MLflow
• Data version control with DVC
• Model serving via FastAPI REST endpoint
• Containerised services with Docker Compose (API, MLflow, Prometheus, Grafana)
• Data drift monitoring with Evidently AI
• CI/CD with GitHub Actions (linting, testing, Docker builds)

🚗 Monocular Depth Estimation

Course: Deep Learning (5DV236)

Real-time obstacle avoidance system: a phone streams its camera to a laptop, which runs YOLO26-Seg detection and Depth Anything 3 monocular metric depth, fuses them into per-object distances, plans a motor command, and drives an ESP32-C3 car over BLE.

• End-to-end fusion server with FastAPI + WebSocket + HTTPS
• Calibrated metric depth via affine fit in disparity space
• Per-instance mask-based distance aggregation (with bbox fallback)
• Phone web app for camera streaming and overlay rendering
• Gradio dev UI and latency benchmarking modes

🔤 Natural Language Processing 1

Course: Natural Language Processing (5DV218 Språkteknologi)

Exploration of fundamental computational linguistics and modern NLP pipelines using spaCy across English and Swedish:

• Rule-based and statistical tokenization, lemmatization, and sentence boundary detection
• Coarse and fine-grained Part-of-Speech (POS) tagging with universal morphological features
• Syntactic dependency parsing and ambiguity visualization ("time flies like an arrow")
• Named Entity Recognition (NER) for standard entity classes
• Swedish language pipeline using sv_core_news_sm for compound words and inflectional morphology
• Vector SVG parse tree visualization with displaCy

📰 Natural Language Processing 2

Course: Natural Language Processing (5DV218 Språkteknologi)

End-to-end text classification benchmarking on the 20 Newsgroups dataset:

• Text preprocessing pipeline with NLTK: case normalization, punctuation filtering, stopword pruning, and Snowball stemming
• Feature representations: Bag-of-Words (CountVectorizer) vs TF-IDF (TfidfVectorizer)
• Supervised classification using Multinomial Naive Bayes, Linear Support Vector Classifier (LinearSVC), and Logistic Regression
• Detailed performance diagnostics: accuracy, macro/micro/weighted F1-scores, precision, recall, and normalized confusion matrix visualizations

📖 Natural Language Processing 3

Course: Natural Language Processing (5DV218 Språkteknologi)

Research project: Implicit state changes in procedural text: probing vs fine-tuning. Investigates whether autoregressive language models implicitly maintain mental representations of world state changes during procedural text processing:

• Benchmarking on the ProPara (science processes) and bAbI Task 2 (two supporting facts) datasets
• Linear layer-wise probing of frozen internal representations across EleutherAI Pythia model scales (160M, 410M, 1B, 1.4B, 2.8B)
• Full causal fine-tuning across models and seeds to evaluate implicit vs explicit state tracking
• Automated statistical analysis with bootstrap confidence intervals
• Complete reproducible experimental pipeline with automated headless Chrome report generation

♟️ Othello

Course: Artificial Intelligence (5DV243)

A competitive Othello (Reversi) game engine featuring:

• Alpha-Beta pruning with iterative deepening search
• Custom position-weighted heuristic evaluation
• Transposition tables and move ordering optimizations
• Time-controlled search within specified limits

🤖 Reinforcement Learning

Course: Reinforcement Learning

Hands-on reinforcement learning project with two parts:

• Multi-armed bandit evaluation against an epsilon-greedy reference baseline
• Multi-agent Pong (PongDuel-v0) using a tabular Q-learning style agent
• Optional rendering, recording, and GIF export for episode analysis

🎰 Spin the Wheel

Course: Software Engineering / Prototype Development

A full-stack "Spin the Wheel" application with:

• Backend: Spring Boot 3.x REST API with OpenAPI documentation
• Frontend: Angular 19 SPA with slot machine UI components
• Firebase integration for data persistence

📊 Statistics

Course: Statistics for Engineers

Statistical analysis investigating the impact of structured micro-breaks on workplace productivity. Includes hypothesis testing, paired t-tests, effect size calculations, and data visualization.

---

🛠️ General Requirements

• Python 3.10+ / 3.11+ (for ML, DL, NLP, Data Privacy, ASP, and Othello projects)
• CUDA GPU (8 GB+ VRAM) (for Monocular Depth Estimation, Deep Learning 3, and NLP 3 Pythia probing/fine-tuning)
• Java 17+ (for Spin the Wheel backend)
• Node.js 18+ (for Angular frontend)
• R 4.x (for Statistics project)
• Clingo 5.x (for ASP project)
• Gym/ma-gym + Pillow (for RL Pong environment and GIF export)
• spaCy & NLTK (with en_core_web_sm and sv_core_news_sm models for NLP 1 & 2)
• SDV / CTGAN & PyTorch (for Data Privacy 3 synthetic data generation)
• Transformers & Accelerate (for NLP 3 LLM probing and fine-tuning)

---

👤 Author

Michail Pettas
Umeå University – AI Masters Program

---

📄 License

These projects are academic assignments. Please respect academic integrity policies if referencing this work.