Topic Machine Learning
Constellation Recognition: GNN + Symbolic AI — Developing a Hybrid Model
We developed a hybrid AI system that detects, separates, and reconstructs Orion, Cassiopeia, and the Big Dipper from noisy, rotated, and scaled point clouds. It combines geometric signatures, an edge-updating GNN, and symbolic graph-isomorphism constraints, reaching a test macro F1 score of 96.91%.
Automatically Generate 3D Models by Inferring Noisy Floor Plans! A Challenge in CubiCasa5K Analysis and Reverse Engineering
Using only floor plan images and SVG annotations—and without reading the official paper or code—the author reverse-engineered CubiCasa5K. The article covers semantic segmentation, random-rotation augmentation, inference of an incomplete sauna door, and conversion of the final result into a 3D model.
AI Economist: Reinforcement Learning in Economics
Experimenting with the AI Economist API to simulate free markets and communism using Reinforcement Learning (RL) agents.
Applying Self Pre-Training Method to GNN for Quantum Chemistry
Explore how self pre-training methods are applied to Graph Neural Networks (GNN) to improve predictions in quantum chemistry.
