Steven Wu (MSR) will present “A Smoothed Analysis of the Greedy Algorithm for the Linear Contextual Bandit Problem” in the Machine Learning and Friends Lunch Thursday Nov. 16 at 11:45 am in CS 150. Abstract and bio follow.

Abstract:

Bandit learning is characterized by the tension between long-term exploration and short-term exploitation. However, as has recently been noted, in settings in which the choices of the learning algorithm correspond to important decisions about individual people (such as criminal recidivism prediction, lending, and sequential drug trials), exploration corresponds to explicitly sacrificing the well-being of one individual for the potential future benefit of others. This raises a fairness concern. In such settings, one might like to run a “greedy” algorithm, which always makes the (myopically) optimal decision for the individuals at hand — but doing this can result in a catastrophic failure to learn. In this paper, we consider the linear contextual bandit problem and revisit the performance of the greedy algorithm. We give a smoothed analysis, showing that even when contexts may be chosen by an adversary, small perturbations of the adversary’s choices suffice for the algorithm to achieve “no regret”, perhaps (depending on the specifics of the setting) with a constant amount of initial training data. This suggests that “generically” (i.e. in slightly perturbed environments), exploration and exploitation need not be in conflict in the linear setting.

Bio:

Steven Wu is currently a Post-Doctoral Researcher at Microsoft Research in New York City, where he is a member of the Machine Learning and Algorithmic Economics groups. He will be joining the Department of Computer Science and Engineering at the University of Minnesota as an Assistant Professor starting in fall 2018. He received his Ph.D. in Computer Science from the University of Pennsylvania in 2017, under the supervision of Michael Kearns and Aaron Roth. His doctoral dissertation “Data Privacy Beyond Differential Privacy” received the 2017 Morris and Dorothy Rubinoff Dissertation Award. His research focuses on algorithm design under different social constraints. In particular, his primary research interest is on data privacy, specifically differential privacy, where he builds tools for data analysis under the constraint of privacy preservation. His recent research also studies algorithmic fairness, especially in the context of machine learning, where he investigates how we can prevent bias and unfairness in algorithmic decision making. He examines problems in these areas using methods and models from machine learning theory, economics, optimization, and beyond.

Mary Farbood (NYU) will present “The Temporal Dynamics of Music Versus Speech Processing” in the Old Chapel Conference Room on Friday, Nov. 17, 2017 at 2:30pm. All are welcome! An abstract is below.

Two studies comparing the temporal dynamics of music and speech are presented. The first focuses on tempo and how it affects key-finding; these results are then compared to various timescales associated with speech processing. The second study examines decoding time of musical structure using a key-finding task and discusses those results in the context of analogous speech research. These experiments highlight both differences and similarities in how music and speech are processed in time.

Lu Wang (Northeastern) will present “What Makes a Good Argument: Understanding and Predicting High Quality Arguments Using NLP Methods” in the Machine Learning and Friends Lunch Thursday Nov. 2 at 11:45 am in CS 150. Abstract and bio follow.

Abstract:

Debate and deliberation play essential roles in politics and civil discourse. While argument content and linguistic style both affect debate outcomes, limited work has been done on studying the interplay between the two. In the first part of this talk, I will present a joint model that estimates the inherent persuasive strengths of different topics, the effects of numerous linguistic features, and the interactions between the two as they affect debate audience. By experimenting with Oxford-style debates, our model predicts audience-adjudicated winners with 74% accuracy, significantly outperforming models based on linguistic features alone. We also find that winning sides employ more strong arguments (as corroborated by human judgment) and debaters all tend to shift topics to stronger ground. The model further allows us to identify the linguistic features associated with strong or weak arguments.

In the second part of my talk, I will present our recent study on retrieving diverse types of supporting arguments from relevant documents for user-specified topics. We find that human writers often use different types of arguments to promote persuasiveness, which can be characterized with different linguistic features. We then show how to leverage argument type to assist the task of supporting argument detection. I will also discuss our follow-up work on automatic argument generation.

Bio:

Lu Wang is an Assistant Professor in College of Computer and Information Science at Northeastern University since 2015. She received her Ph.D. in Computer Science from Cornell University and her bachelor degrees in Intelligence Science and Technology and Economics from Peking University. Her research mainly focuses on designing machine learning algorithms and statistical models for natural language processing (NLP) tasks, including abstractive text summarization, language generation, argumentation mining, information extraction, and their applications in interdisciplinary subjects (e.g., computational social science). Lu and her collaborators received an outstanding short paper award at ACL 2017 and a best paper nomination award at SIGDIAL 2012. Her group’s work is funded by National Science Foundation (NSF), Intelligence Advanced Research Projects Activity (IARPA), and several industry gifts (Toutiao AI Lab, and NVIDIA GPU program). More information about her research can be found at www.ccs.neu.edu/home/luwang/.

Ana Francisco will present “Audiovisual processing in dyslexia” in the Cognitive Bag Lunch series, Wednesday at noon in Tobin 521B. Abstract below. All are welcome!

The process of learning to read is lengthy and cognitively demanding. Nevertheless, the majority of us, if properly instructed, learn to read without problems. There is a significant minority of individuals, though, who struggle to acquire this fundamental skill. At least in alphabetic orthographies, a crucial phase in developing the ability to read is learning the socially agreed-upon associations between letters and speech sounds. In this sense, learning to read ultimately relies on the formation of automatic audiovisual associations. Importantly, this reliance of reading on audiovisual associations persists into adulthood, such that the expert reader continues to depend on these audiovisual objects. Yet, the relationship between reading ability and audiovisual processing is still not fully understood. In this talk, I will focus on such a relationship and argue, showing behavioral and neuroimaging evidence, for the presence of an audiovisual processing deficit in dyslexia.

Beomjoon Kim (MIT) will present “Learning to Guide Task and Motion Planning by Predicting Constraints” in the Machine Learning and Friends Lunch Thursday Nov. 2 at 11:45 am in CS 150. Abstract and bio follow.

Abstract:

In robotics, it is essential to be able to plan efficiently in high-dimensional continuous state-action spaces for long horizons. For such complex planning problems, unguided uniform sampling of actions until a path to a goal is found is hopelessly inefficient, and gradient-based approaches often fall short when the optimization manifold of a given problem is not smooth. In this talk, I will introduce two of our recent approaches for using past planning experience to learn to predict constraints for guiding a planner. In the first part, I will introduce a technique that uses generative adversarial networks and importance sampling estimation to learn an action distribution that restricts the search space to a promising region, when an input is represented with a vector. In the second part, I will discuss the limitation of a vector representation in complex robot planning settings, and propose a more suitable representation for guiding a search, called a score-space representation. With this representation, we can predict a constraint on the search space by optimizing a black-box function. We empirically show that a planner is able to find a solution more efficiently using these approaches on a various robot task and motion planning problems.

Bio:

Beomjoon Kim is a PhD student at MIT CSAIL under the supervision of Leslie Pack Kaelbling and Tomas Lozano-Perez. His recent research focuses on developing machine learning algorithms for complex robot planning problems, in which problems involve reasoning about both discrete, logical structures and continuous, geometric structures of the world. In the past, he has worked on robot learning from demonstrations and reinforcement learning. He received his MS.c from McGill University under the supervision of Joelle Pineau, and received BMath from University of Waterloo.

Andrea Cataldo will present in the Cognitive Bag Lunch series, Wednesday at noon in Tobin 521B. All are welcome!

Danny Fox of MIT will present “Exhaustivity as cell identification” in the Linguistics Colloquium series organized by the GLSA. The talk is in ILC N400 at 3:30 – all are welcome! The abstract is below.

Abstract:

Under the Grammatical Theory of Scalar Implicatures (GT of SIs), SIs are logical entailments of ambiguous sentences – entailments of grammatical representations containing a (covert) focus sensitive operator, exh. Conceptual and empirical arguments have been presented in favor of GT. GT is consistent with conversational maxims that can be defended on a-priori grounds. By contrast, the pragmatic alternatives require the rejection of what are arguably basic truisms. In additions, GT is supported on various empirical grounds, among them considerations of modularity/blindness, the interaction between ignorance inferences and SIs, embedded implicatures, and various other areas of grammar in which Exhaustification plays a role (for a review and relevant references, see Chierchia, Fox and Spector 2012, Fox 2014, and Chierchia 2017).

However, the definitions of exh provided in most versions of GT have been suspiciously close to what is derived by Neo-Gricean mechanisms. I will argue that this conceptual difficulty might be eliminated when we look at the presuppositions of questions. Building on and modifying work of Dayal (1996) I will argue that exh is involved in the statement of these presuppositions. Specifically: a question, Q, presupposes that every cell in the partition it induces can be identified by a member of Q, via Exhaustification (?C?Partition(Q)[?p?Q[Exh(Q,p)=C]]), and conversely, that every member of Q identifies a cell in the partition (?p?Q[?C?Partition(Q)[Exh(Q,p)=C]]).

Drawing on recent work with Moshe Bar-Lev, I will provide an alternative definition of exh that is quite distinct from operators that reflect Neo-Gricean mechanisms. I will also argue that this definition can be understood based on considerations that come from the role of exh in question semantics (its role in cell identification). The empirical arguments will be based primarily on the distribution of negative islands and “mention some” readings.

Kyunghyun Cho (NYU) will present “Deep Learning, Where are you going?” in the Machine Learning and Friends Lunch Thursday Oct. 20 at 10:00am in CS 150. Abstract and bio follow.

Abstract:

There are three axes along which advances in machine learning and deep learning happen. They are (1) network architectures, (2) learning algorithms and (3) spatio-temporal abstraction. In this talk, I will describe a set of research topics I’ve pursued in each of these axes. For network architectures, I will describe how recurrent neural networks, which were largely forgotten during 90s and early 2000s, have evolved over time and have finally become a de facto standard in machine translation. I continue on to discussing various learning paradigms, how they related to each other, and how they are combined in order to build a strong learning system. Along this line, I briefly discuss my latest research on designing a query-efficient imitation learning algorithm for autonomous driving. Lastly, I present my view on what it means to be a higher-level learning system. Under this view each and every end-to-end trainable neural network serves as a module, regardless of how they were trained, and interacts with each other in order to solve a higher-level task. I will describe my latest research on trainable decoding algorithm as a first step toward building such a framework.
Bio:

Kyunghyun Cho is an assistant professor of computer science and data science at New York University. He was a postdoctoral fellow at University of Montreal until summer 2015, and received PhD and MSc degrees from Aalto University early 2014. He tries best to find a balance among machine learning, natural language processing and life, but often fails to do so.

David Kellen (Syracuse University) will present in the Cognitive Bag Lunch series, Wednesday at noon in Tobin 521B. All are welcome!

The Cognitive Science Society is seeking applications for a new Graduate Student Representative to the society’s Governing Board. The GSR will be a current graduate student in a cognitive science related field who will represent student interests as a non-voting member of the Governing Board. In addition, the GSR will sit on the society’s Communications Committee and will manage the twitter and facebook accounts of the society, staying active in updating these on a regular basis with society news and publications in the society journals. We anticipate a total commitment of 1-3 hrs/week plus a few more significant efforts (e.g., before Board meetings or in a revision of the website).

This position carries a two year term and has as compensation: society membership, free registration at the 2018 and 2019 meetings, travel reimbursement for the meetings each year (up to $1000 for domestic and $2000 for international), and a $1000 stipend each year. Applicants should have the expectation to be in a graduate program during their full term (e.g., should be relatively early on in their graduate career).

Applications should be sent to mcfrank@stanford.edu with the title “CogSci GSR” and should include a CV and a brief (200-300 word) statement describing the candidate’s professional goals, interest in the position (for example, what would be priorities as a representative), and relevant experience in communication and/or social media. Evidence of prior commitment to the Cognitive Science Society (e.g., membership, annual meeting attendance) will be taken into account in the process but is not necessary. The applicant should also arrange for a brief letter of recommendation from an academic mentor (email or letterhead) to be sent separately to the same address.

Please submit applications by 11/6 for full consideration by the Communications Committee. The new GSR will be notified in late November and the current term will start on 12/1.