After a few months of work, I've released the first stable version of a filesystem benchmark that is able to simulate the aging of filesystems and can be used to test filesystems for their performance degradation over time. The project is open-sourced thanks to Red Hat's open source policy
I've trained a ViT model capable of nowcasting weather radar images for several hours in future. The model uses a visual transformer to predict the weather based on satellite images. Along with the model, I've also open-sourced a tool for easy training and deployment, and a script for global weather prediction. Please see the GitHub repository for more information.
Submitted a paper to the MEDICON 2023 conference as the first author. Developed a deep learning model based on the V-Net architecture for automatic inference of the radiological planes of the heart from 3D MRI sequences. This model improves the accuracy and efficiency of medical imaging by generating heatmaps of probable plane positions, enabling faster and more effective diagnosis and treatment. Demonstrated the network's capability to accurately locate cardiac landmarks, even with anisotropic voxel data.
As a school project we have been tasked to train a Generative Adversarial Network on fingerprint images. Trained network should be later used to generate synthetic fingerprint images and break the fingerprint recognition system. The project was done in Python and the neural network was implemented in PyTorch. We have been able to combine trained network with genetic algorithm to generate such synthetic fingerprint image which was able to break the fingerprint recognition system.
As a school project we have been tasked to train a Dense passage retrieval on open domain question answering. Even though it was only a school project, I have invested a lot of time and effort into it, to learn as much as possible. The project was done in Python and the neural network was implemented in PyTorch. I've based the implementation and training on the original Facebook AI Research paper. I've conducted dimensionality reduction in order to discover the ability of the model to perform on a smaller vector space. The project was a great learning experience and I have been able to produce a interesting results.
As a school project we have been tasked to train a OCR neural network. Since it was only a school project, we haven't aspired to create a perfect OCR, but rather build the best we could in the given time frame. The project was done in Python and the neural network was implemented in PyTorch. We have used relatively small custom convolutional architecture, optimized for CTC loss. As dataset we used Brno Mobile OCR Dataset. Overall, the project was a great learning experience and we have been able to achieve a decent OCR accuracy.