The Bloomberg Data Science Research Grant Program aims to support cutting-edge research in the broad field of machine learning, including specific areas such as natural language processing, information retrieval, machine-translation and deep neural networks. In April 2015 we announced our first round of recipients and in October 2015 we announced our second. Today, we are pleased to announce the winners of our third round of awards.
Out of hundreds of applications from faculty members at universities around the world, a committee of Bloomberg researchers selected the proposals of eight research projects:
Daniel Hsu (Columbia University) and Kamalika Chaudhry (University of California, San Diego):
Spectral Learning with Prior Information with Applications to Topic Models:Complex statistical models are challenging to fit to large, high-dimensional data sets. Although several recent developments in machine learning have led to scalable fitting methods based on simple algebraic techniques, they are unable to incorporate prior knowledge constraints into the model fitting. Professors Chaudhuri and Hsu will develop new extensions of these scalable methods that can handle such constraints, and they will apply these methods to perform comparative analyses of large document corpora.
Yisong Yue (California Institute of Technology):
Dynamic Interpretability in Machine Learning: In many cases, we desire predictive models that are both accurate and interpretable. However, the goals of accuracy and interpretability are often in conflict with each other, since interpretable models tend to be overly simple. In this project, Professor Yue proposes to study an alternative notion of interpretability, which he calls “dynamic interpretability”. The goal of dynamically interpretable models is to make predictions that are interpretable, rather than have the model itself be explicitly interpretable. With this alternative goal, one can circumvent much of the inherent tension between accuracy and traditional “static” interpretability, and move one step closer to interpretable production-strength models.
Shay Cohen (University of Edinburg):
Latent-Variable Spectral Learning Kernelization for NLP: The natural language processing community is now seeking to find succinct representations of text that can be geometrically interpreted. One approach to do so is through the use of so-called spectral learning, which makes use of linear-algebraic techniques to create such representations. Spectral methods have some limitations at the moment with respect to the level of expressivity they can bundle in geometric representations. To address these issues, Dr. Cohen intends to further develop spectral methods so that they can elicit richer representations than before.
Stephen Becker (University of Colorado at Boulder):
Online clustering of time-sensitive data: Modern datasets, such as news articles or tweets, have a time-sensitive nature which is not well captured by traditional offline supervised or unsupervised learning. Prof. Becker’s work will extend some clustering algorithms to work with streaming data and adapt modern variance-reduction strategies to an online setting.
Christopher Manning (Stanford University):
Character-level neural network sequence models for varied text named entity recognition: Finding references to people and organizations is key to computers understanding human language text, but simple techniques like name lists often fail when informal names like Tay, Fif or W are used. Prof. Manning’s proposal will attempt to improve the state-of-the-art by exploiting the flexibility and power of character sequence level deep learning models.
Noah Smith (University of Washington), Amber Boydstun (University of California, Davis), Philip Resnik (University of Maryland), Justin Gross (University of Massachusetts, Amherst):
What’s The Angle? Disentangling Perspectives from Content in the News: News coverage doesn’t just involve the topic or issue being covered; perspectives are also woven into the language used to discuss an issue. Perspective is connected with a story’s angle—the aspect of the story that the journalist chooses to highlight and develop. This project develops computational models that use linguistic signals found in text to uncover patterns of perspective in news sources, disentangling perspective from content.
Mohit Bansal (TTI-Chicago):
Multimodal Event Summarization: Prof. Bansal proposes a multimodal event highlighting and summarization system that uses structured neural models to automatically select the most important events in multimodal, free-form text and video databases and verbally summarize them. He will focus on applications of such a model in automatic news highlighting and video surveillance alerting.
Maarten de Rijke (University of Amsterdam):
Contextual Entity Recommendation: Entities are at the heart of many information needs. Contextual entity recommendation methods help users understand and navigate around complex information spaces by recommending entities in the context of a search task. Prof. de Rijke proposes a series of methods based on deep learning entity recommendation. How can such methods scale to very large sets of entities? How can they use both text and knowledge bases to arrive at entity recommendations
Congratulations to the winners. If you are interested in applying for future grants through The Bloomberg Data Science Research Grant Program, click here.