Implementation of an ImageGPT-like model designed to generate images from a given context. The system uses a VQ-VAE to learn a discrete codebook of visual tokens, which are then modeled by a decoder-only Transformer (GPT-like architecture) to generate coherent sequences of tokens corresponding to images.
An Information Retrieval (IR) system was developed in Python and evaluated against a baseline implementation using PyTerrier. The system adopts an efficiency-oriented retrieval strategy based on the DAAT approach, further optimized with the WAND algorithm to reduce the number of scored documents. The inverted index is compressed using the Elias Fano encoding scheme, improving memory efficiency while maintaining fast access. Experiments conducted on the MS MARCO dataset show that the system achieves a median query execution time of 0.041 seconds for retrieving the top 10 documents.
The application is a mobile app developed in Kotlin that connects to an EEG device via Wi-Fi. It allows users to monitor their mental workload in real time through a wearable EEG sensor. The system records brain activity data and uses a machine learning classifier to estimate the user's level of fatigue. The model is fine-tuned with personal user data and stored locally on the device to preserve privacy. Fatigue detection relies on the analysis of theta and alpha frequency bands, which are indicative of cognitive load and tiredness. When increased fatigue is detected, the application notifies the user, suggesting a break.