Explaining Neural Networks with Layerwise Relevance Propagation (LRP)
The project was focused on the problem of explaining Neural Network Decisions with LRP method. Students analyzed interpretations of the CNN-based classification and aggregated results for the MNIST dataset. Also, students applied LRP to the Fashion MNIST dataset from Zalando and presented observed results in the presentation.
You can try out the Demo here.
Interactive Named Entity Recognition
The goal of the TraiNER (Train Named Entity Recognition) project was to develop a system that illustrates an Active Learning Process in the field of NER. For this purpose, a system consisting of several components was implemented.
It allows the user to train a NER model by marking spans of text in an textmarker-like fashion.
The frontend allows an user to upload documents and entities, whereupon a NER model is trained. The backend persists every step of the way. The machine learning component of the project is iteratively training a model, initially based on a string match. A feedback loop with the frontend allows the user to improve or approve sampled sentences, whereupon the model is further trained. An overview of the results can be found in the presentation.