The following markdown is an overview my most important data science projects. All deep-learning models have been implemented in PyTorch. The more classical machine learning algorithms have been implemented using Scikit-learn. The overview is split into three sections, based on the topic of each project.

Description: This project was my master thesis. The thesis is done in collaboration with ClimbAlong at NorthTech ApS, who provided an annotated dataset and a pretrained Mask R-CNN for pose estimation. The task of the project is to extend the provided Mask R-CNN, such that the pose estimator leverages the temporal information of a video to improve the performance of the model. The project experiments with various architectures (3D-conv, bi-convLSTM and transformer), such that we ended up with the best setup.

Keywords: Mask R-CNN, bi-ConvLSTM, Transformer

Description: This project was my bachelor thesis, where I implemented a Stacked hourglass. This was then followed by an interpolation of the model, which was done by interpreting the effects of certain parts of the model, as well as the structure of the latent space. Based on these interpretations I concluded, that there were some redundancy and misplaced samples in the latent space of the model. Lastly, I used the obtained information about the model to improve the performance and remove redundancy of the model. This was done by modifying the architecture to include an autoencoder with a reduced information bottleneck.

Keywords: Stacked hourglass, Autoencoder, XAI

Description: This project aimed at explaining U-net. In the project I implemented and trained U-Net on a chest X-ray dataset. I then interpreted the model, where I concluded, that (1) the model was affected by the bias in the training data, due to an imbalanced class distribution, (2) the job of the encoder, bottleneck and the convolutions of the decoder, and (3) the most important features of an input-image.

Keywords: U-Net, Segmentation, Medical Image Analysis, XAI

Description: This project aimed at implementing multiple question answering models. This includes models that can classify whether a text contains the answer to a question, and also models that can extract said answer. Additionally, we analyse the dataset via these models to interpret how it works. All of the models ended up succeeding in both the classification and token classification tasks

Keywords: Multilingual, Question answering, XLM-RoBERTa

Description: This projected was centered around the paper Generalization without Systematicity: On the Compositional Skills of Sequence-to-Sequence Recurrent Networks. The paper focused on translating navigation commands to a sequence of actions. The aim of the project was to (1) replicate the results of the paper, and (2) extend the paper, such that the experiements were done with a transformer-based model instead.

Keywords: Machine translation, LSTM, GRU, Transformer

Description: This project aimed at implementing and evaluating search algorithms for news articles. This included (1) implementing and evaluating various indices, (2) implementing and evaluating various ranking models, and (3) present, analyse and discuss the results.

Keywords: Information retrieval, GloVe, BM25, BERT, T5