The thesis focused on Learnable Dynamic Voltage Optimization in radio units at Ericsson. It combined classification and regression models with traffic-aware thresholds to reduce power consumption while maintaining performance. The project also involved explainability using Shapley Additive Explanations (SHAP).
The thesis focused on making board games more accessible for people with visual impairment by converting physical cards into a digital hand via camera + image recognition. It was tested on Catan and achieved high card recognition accuracy.
This project builds on WACO by integrating meta-learning to reduce re-training time when adapting to new hardware. With just 300 samples, it outperformed the original WACO on a new machine.
Transfer learning with ConvNeXT. Accuracy 0.88 on held-out test set.
RAG pipeline using Pinecone vector DB. Best config: e5-base + Mistral, beating GPT-3.5-turbo on this dataset.
Financial QA using ARKMan, hybrid search, chain-of-thought, and query decomposition; FinQA dataset for evaluation.
3000 docs from 20 Newsgroups. Best coherence with 10 topics and α = β = 0.1.
CNN for 12 species from bark images. Ablation study highlights: residuals, normalization, batch norm. Added OOD vs ID task (100% on that split).
11-node grid model minimizing generation cost while meeting demand (active/reactive). Best found cost: 186.29 SEK (non-convex; not guaranteed global optimum).
Comparing LoRA with feature extraction for task adaptation. LoRA achieved Acc/F1 ≈ 0.93 vs 0.81.
Best baseline: Logistic Regression + TF-IDF (Acc 0.89) on annotated social media comments.
MCTS with UCT. Win rate: 0.86 vs random; 0.83 vs blocking strategy.
Hybrid preferences using prior ratings, genres, and similar users.
Deterministic bigram per category achieved highest semantic similarity (0.56) and strong BLEU on this dataset.
IBM Model 1 on parallel EN-SV-DE-FR corpus.