who: Maja Rudolph – Columbia
when: 11:45 A.M. 1:15 P.M., Thursday, March 29th
where: Computer Science Building Rm 150
food: Athena’s pizza
“Exponential Family Embeddings”
Abstract: Word embeddings are a powerful approach for capturing semantic similarity among terms in a vocabulary. Exponential family embeddings extend the idea of word embeddings to other types of high-dimensional data such as count data from a recommendation system or real-valued data from neural recordings. Exponential family embeddings have three ingredients; embeddings as latent variables, a predefined conditioning set for each observation called the context, and a conditional likelihood from the exponential family. The embeddings are inferred with a scalable algorithm based on stochastic gradient descent. In this talk, I discuss three highlights of the exponential family embeddings model class: (A) The approximations used for existing methods such as word2vec can be understood as a biased stochastic gradients procedure on a specific type of exponential family embedding model. (B) By choosing different likelihoods from the exponential family we can generalize the task of learning distributed representations to different application domains. (C) Finally, the probabilistic modeling perspective allows us to incorporate structure and domain knowledge in the latent space. With dynamic embeddings, we can study how word usage changes over time and structured embeddings allow us to learn embeddings that vary across related groups of data. Key to the success of our method is that the groups share statistical information and we develop three sharing strategies: dynamic modeling, hierarchical modeling, and amortization.
|Bio: As a computer science PhD student at Columbia University, Maja Rudolph studies probabilistic modeling and approximate inference. Together with her advisor David Blei, she works on embedding models and explores how they can be used to find rich, interpretable structure in large data sets. In 2013, she obtained a BS in mathematics from MIT.
Robert Siegler (Carnegie Mellon University http://www.psy.cmu.edu/~siegler/), will speak on “Numerical Development” from 1-2pm in ILC room N400. The talk is sponsored by the Developmental Science Initiative. An abstract follows.
Abstract: In this talk, I attempt to integrate two crucial aspects of numerical development: learning the magnitudes of individual numbers and learning arithmetic. Numerical magnitude development involves gaining increasingly precise knowledge of increasing ranges and types of numbers: from nonsymbolic to small symbolic numbers, from smaller to larger whole numbers, and from whole to rational numbers. One reason why this development is important is that precision of numerical magnitude knowledge is correlated with, predictive of, and causally related to both whole and rational number arithmetic. Rational number arithmetic, however, also poses challenges beyond understanding the magnitudes of the individual numbers. Some of these challenges are inherent; they are present for all learners. Other challenges are culturally contingent; they vary from country to country and classroom to classroom. Our findings indicate that a largely ignored culturally contingent variable, distributions of problems in mathematics textbooks, substantially influences learning of rational number arithmetic. Generating theories and data that help children surmount the challenges of rational number arithmetic is an important goal for numerical development research.
Colin Phillips (University of Maryland, https://www.colinphillips.net) is presenting “Speaking, understanding, and grammar” in the Department of Linguistics on Friday, March 30th, at 3:30PM, in the Integrative Learning Center in room N400. The abstract is below.
Abstract: We speak and understand the same language, but it’s generally assumed that language production and comprehension are subserved by separate cognitive systems. So they must presumably draw on a third, task-neutral cognitive system (“grammar”). For this reason, comprehension-production differences are a thorn in the side of anybody who might want to collapse grammar and language processing mechanisms (i.e., me!). In this talk I will explore two linguistic domains from the perspective of comprehension and production. In the case of syntactic categories, I will show that the same underlying mechanisms can have rather different surface effects in comprehension and production. In the case of argument role information, I will show an apparent conflict between comprehension and production. In production, argument role information tightly governs the time course of speech planning. But in comprehension, initial prediction mechanisms seem to be blind to argument role information. I argue that both the similarities and contrasts can be captured under a view in which the same cognitive architecture is accessed based on different information, i.e., sounds for comprehension, messages for production. I will discuss the relation between this and other ways of thinking about comprehension-production relations, drawing on a combination of behavioral and electrophysiological evidence.
The Music and Language CogSci Incubator (https://blogs.umass.edu/cogsci/2018/03/10/music-and-language-cogsci-incubator-with-acevedo-and-temperley/) will begin on Tues. April 3rd at 2:30 in ILC N458 with a discussion of David Temperley’s The Musical Language of Rock. Participants are encouraged to bring questions and discussion points that don’t assume everyone has read the book (in other words, please come even if you haven’t even cracked its spine). The book turns out to assume a fair bit of music theory, so clarification questions are very much appropriate (and we might pick something else to discuss the next week). If you are having trouble getting the book on time (OUP ships quickly), please contact Joe Pater.
The fourth annual UMass CogSci Workshop will be held in conjunction with Sue Carey’s visit to the campus on April 20th (https://blogs.umass.edu/cogsci/2018/01/30/sue-carey-friday-april-20th-at-330/). The workshop will consist of a poster session from 2:15-3:15; please submit your poster info here: https://goo.gl/forms/mt9v6NYU30zywGNe2. As always, previously presented work is allowed, even encouraged (don’t print a new poster if you can use one you already have!).
Ibrahim Dahlstrom-Hakki (Landmark College) will speak on “Teaching Students with Disabilities Online: Language-Based Challenges and Cognitive Access” in the next cognitive brown bag, Wednesday, March 28 in Tobin 521B from 12-1:20. The abstract is below.
Many students with Learning Disabilities (LD), Attention Deficit Hyperactivity Disorder (ADHD), and Autism Spectrum Disorder (ASD) struggle in online learning environments. Online courses tend to place high demands on their language processing and executive function skills. In this NSF funded study (DRL-1420198), we look at some of the barriers facing students with disabilities learning statistics concepts through online discussions. We report on the impact of a Social Presence manipulation on their performance and some of the language-based difficulties involved in assessing their knowledge.
There are lots of people coming to UMass (and Smith) to give CogSci talks after break – here is a list.
March 19th, 4pm Weinstein Auditorium, Smith College
Richard Aslin (Yale University)
“What Looking Times and Brain Imaging Can Tell Us About Attention and Learning in Infants”
March 30th, 3:30, ILC N400
Colin Phillips (https://www.colinphillips.net) Linguistics colloquium
April 13th, 2:30, ILC N400
Music and Language CogSci Incubator
Stephanie Acevedo (https://stefanieacevedo.com)
David Temperley (http://davidtemperley.com)
April 19th, 1:00, ILC N400
Robert Siegler (http://www.psy.cmu.edu/~siegler/) “Numerical Development”
April 20th, 3:30, ILC S131
Sue Carey (https://psychology.fas.harvard.edu/people/susan-e-carey) “Do Non-Linguistic Creatures have a Fodorian (Logic-Like/Language-Like) Language of Thought?”
April 27th, 3:30, ILC N400
Meghan Clayards (https://www.mcgill.ca/linguistics/people/faculty/clayards) Linguistics colloquium
May 19-20th, South College Room W245
Lynschrift18: A conference to honor Lyn Frazier on the occasion of her retirement (https://blogs.umass.edu/lynschrift18/program/)
In the next cognitive brown bag, 3/21, 12:00-1:20, Tobin 521B, Quili Ma (PBS) will present: “Testing recognition models with joint single-item and forced-choice recognition.” The abstract follows.
Two models of recognition memory: the double-high threshold (2HT) and the unequal variance signal detection (UVSD) models were compared with the forced-choice experiment. Participants studied a list of words. In the test phase, they first responded “studied” or “not studied” to single word recognition for a few trials and then indicated the “studied” word in 2AFC trials. A 2AFC word pair contained a target and a lure which had both been previously given the same response. So one of the words was incorrectly recognized. The 2HT model predicted that forced-choice trials’ accuracy should increase from the single-recognition trials because between the two words, one had no memory evidence retrieved and the other word was very likely to be a detection result which had infallible information. On the contrary, the UV model predicted lower accuracy of the forced-choice trials because the target had to be compared with a lure which had very strong evidence of being studied. The empirical data turned out to support the UV model’s prediction.
Richard Aslin (Yale University), 4pm, Monday March 19th, Weinstein Auditorium, Smith College [co-sponsored by the Developmental Science Initiative and the Office of the President, Smith College]
“What Looking Times and Brain Imaging Can Tell Us About Attention and Learning in Infants”
During the course of development, human infants gather information about the external world without the benefit of an extensive base of knowledge that adults automatically bring to bear on perceptual, motor, cognitive, and language tasks. What mechanisms allow infants to acquire this initial level of information and how does that information guide subsequent learning? Clearly, most learning that occurs in infancy, and a substantial amount of learning in adulthood, is performed without instruction—it is implicit and based on an analysis of the distributional properties of environmental stimulation. In this lecture, Prof. Aslin will present research findings and implications, from studies using behavioral and brain optical imaging techniques.
After 33 years on the faculty in the Department of Psychology and then Brain and Cognitive Sciences at the University of Rochester, Richard Aslin has joined Haskins Laboratories as a senior scientist. He will be re-establishing a BabyLab to carry on the outstanding tradition of developmental research at Haskins, complementing the on-going studies of older infants and young children by other Haskins scientists.