My current and past research projects

Here is an overview of projects I am working on. If you have questions or want to discuss anything, please send me an email at my_last_name (at) seas.harvard.edu

 

LSTMVis:  Visual Analysis for Recurrent Neural Networks

The usual neural network has millions of learned parameters. While these parameters lead to great results, there are too many for a human to make sense of. As good as our understanding for the underlying algorithms are, we cannot quite understand what a network is learning. Thus, we need assistance in the debugging process to understand what the network learned and where it goes wrong. 

Our vision for this project is to build an interactive tool that can be used to explore what a network has learned. This tool can be used to find flaws in the data, in the architecture and to find interesting new patterns in your data. The tool can be found on our website and the paper can be found here.

 

Comparison of traditional and Deep Learning for Patient Phenotyping

In secondary analysis of health records, most studies neglect the valuable information in written parts of the EHR such as discharge notes or nursing notes. Studies suggest that up to 80% of information within EHR's can be found within the text. Therefore, for accurate patient phenotyping, it is crucial to consider the texts as well. The reason why most studies do not bother to include textual information is because even the most advances text mining systems require a heavy supervision from clinicians. The clinicians not only have to build a very task-specific labeled dataset, they also have to define phrases associated with the desired phenotype to build a rule-based named entity extraction pipeline. 

In this project, we are building a classificatio system based on a convolutional neural net that learns the phrases that are associated with a patient phenotype. This makes it very applicable to different tasks and reduces the workload for the involved physicians. Preliminary tests have shown very promising results and we are planning submit a paper soon. Stay tuned! 

 

Help identifying relevant references - Writer's Aid 2.0

Researchers spend a lot of time trying to find references relevant to them. They tend to look for papers in venues known to them, even though many fields in science and engineering are of a collaborative nature. One reason for this is that researchers know the relevant keywords they have to look for in these venues and fields. Other fields might have different terminologies or the researchers might simply not be aware that the same problem is being attacked in different fields. 

We aim to first build a system that identifies relevant resources and then recommends them. You will never have to manually look for references again, and can focus on what researchers love - writing! My results so far have shown that a system that predicts paper based on a citation context outperforms traditional information retrieval and will implement my predictive system as a browser add-on soon. 

 

Automated Narrative Facilitation Between Two Conflicting Parties 

Research in psychology has shown that collaborative development of a narrative helps to bring two parties with different views closer together. In this case, the resulting story is less important than the actual process of how to come to a shared story. However, mediators are required to speak both languages (in case of the two parties speaking different languages) and are often seen as biased if they are of one of the two nationalities. During this project, we are focusing on the conflict between Israel and Palestine. Two adolescents have to take turns in developing a comic and have full control over backgrounds, characters, speech bubbles and other objects in the story. I developed a facilitation algorithm that leverages a theory in psychology that states that a good shared narrative has to have a clear escalation and de-escalation. The algorithm can measure the current escalation of a story based on sentiments of individual parts and sends notification to the individual interfaces, nudging the people to talk about their problems or to propose solutions. 

This work is done as part of the Per.Te project at the Bruno Kessler Foundation in Trento, Italy under the supervision of Massimo Zancanaro and Oliviero Stock.