This tutorial summarizes the literature on knowledge extraction in scientific domains, such as biomedical texts.
This tutorial discusses the machine learning techniques popular in the information retrieval subcommunity such as ranking techniques, pagerank, etc.
This tutorial gives an introduction to the control theory, in particular, discussing the trajectory optimization techniques.
This tutorial discusses probabilistic models popular in the NLP lexical semantics community.
This tutorial discusses probabilistic models for social and other network data.
This tutorial presents a brief survey of active learning, submodular functions, and the interesting algorithms and analyses at their intersection. Minimal background knowledge is assumed, and emphasis is placed on open problems and gaps between theory and practice. Slides at: http://ml.cs.washington.edu/www/media/presentations/submodularity_tutori...
Our world is becoming more data driven. With the spread of ubiquitous sensors, network connectivity, and massive storage capabilities, we are able to collect more and more data. But our computation and analysis capabilities have not increased at a comparable rate. Computer scientists are facing looming questions such as "How do we deal with the massive amounts of data we are collecting? How can we extract value out of data?" A sub-question relevant to machine learning researchers is "what role will machine learning and data mining play?" Through a survey of current sources of Big Data and analysis workflow patterns, this talk aims to shed light on the latter question.
We describe a relaxed and generalized notion of maximum entropy problems for multinomial distributions. By introducing a simple re-parametrization we are able to derive an efficient homotopy tracking scheme for the entire relaxation path using linear space and quadratic time. We also show that the Legendre dual of the relaxed maximum entropy problem is the task of finding the maximum likelihood estimator for an exponential distribution with L1 regularization. Hence, our solution can be used for problems such as language modeling with sparse parameter representation.
Recent work has demonstrated effective learning algorithms for a variety of semantic parsing problems, where the goal is to automatically recover the underlying meaning of input sentences. Although these algorithms can work well, there is still a large cost in annotating data and gathering other language-specific resources for each new application. In this talk, I will describe efforts to address these challenges by developing scalable, probabilistic CCG grammar induction algorithms. I will present recent work on methods that incorporate new notions of lexical generalization, thereby enabling effective learning for a variety of different natural languages and formal meaning representations. I will also describe a new approach for learning semantic parsers from conversational data, which does not require any manual annotation of sentence meaning. Finally, I will sketch future directions, including our recurring focus on building scalable learning techniques while attempting to minimize the application-specific engineering effort. Joint work with Yoav Artzi, Tom Kwiatkowski, Sharon Goldwater, and Mark Steedman
Segmental Conditional Random Fields (SCRFs) are a mathematical model in which an observation sequence is segmented into labeled chunks. For example, an audio sequence may be segmented into words, or a video segmented into scenes. These models have two key characteristics. First, in contrast to frame-based systems, long-span features can be used - features which relate an entire span of observations to a label, and which may make reference to precisely defined segment boundaries. Second, the models are log-linear in form, and allow for the convenient integration of features derived from heterogeneous information sources. This talk will provide an overview of SCRFs, and in particular their use in speech recognition. First, we describe the basic model. Then, we present results in which SCRFs are used to add new information sources to improve the performance of a state-of-the-art system. Finally, we present initial results in which the framework is used to do recognition from the ground up.