Rosie Cowell of UMass will present “Two surprising results predicted by a representational account of memory:
normal aging shields us from semantic interference, and we can do cued recall without engaging the hippocampus” at noon Wednesday Sept. 13 in Tobin 521B. All are welcome!

who: Scott W. Linderman, Columbia University
when: noon, Thursday, May 4
where: Computer Science Building Rm 150
food: Antonio’s pizza

“Bayesian Learning and Inference in Recurrent Switching Linear Dynamical Systems”

Abstract:  Many natural systems, such as neurons firing in the brain or basketball teams traversing a court, give rise to time series data with complex, nonlinear dynamics. We can gain insight into these systems by decomposing the data into segments that are each explained by simpler dynamic units. I will present a model class that builds on the switching linear dynamical system (SLDS), leveraging its combination of discrete and continuous latent states to discover dynamical units. Our recurrent SLDS will go one step further: by learning how transition probabilities depend on observations or continuous latent states, we will better explain switching behavior. Our key innovation is to design these recurrent SLDS models to enable Pólya-gamma auxiliary variable techniques and thus make approximate Bayesian learning and inference in these models easy, fast, and scalable.

Bio :  Scott is a postdoctoral fellow in the labs of Liam Paninski and David Blei at Columbia University. He completed his Ph.D. in Computer Science at Harvard University under the supervision of Ryan Adams and Leslie Valiant, and he received his B.S. in Electrical and Computer Engineering from Cornell University. Prior to graduate school, he worked at Microsoft as a software engineer on the Windows networking stack. His research is focused on machine learning, computational neuroscience, and the general question of how computer science and statistics can help us decipher

Gaja Jarosz of Linguistics is offering a new course in Computational Cognitive Modeling (Ling 692C Fall 2017, MW 4-5:15). She very much welcomes participants from other departments, including advanced undergraduates with the requisite background (see prerequisites below). Anyone interested should contact Prof. Jarosz directly.

Description. This is an advanced introduction to computational linguistics intended to complement CS 585 Natural Language Processing, focusing on computational approaches applied to cognitive and linguistic questions. The focus of the course will be on computational implementations of linguistic theories and on models of behavioral linguistic data, such as data from first language acquisition, (human) language processing experiments, and/or categorization/perception tasks. We will consider cognitive computational models spanning various domains in linguistics from phonetics to semantics.

Prerequisites:  Python programming experience (Ling 492 or similar)

There will be two First Year Project Reports, by Jeffery Durbin and Dimitris Stergiedis, in the Cognitive Bag Lunch on Wednesday May 3 at 12pm in Tobin 521B. Titles and abstracts below. All are welcome!

Presenter: Jeffery Durbin

Tile: Characterizing the Visuospatial Sketchpad: Rehearsal and Retrieval in Visual Short-Term Memory

Abstract: Sternberg’s (1966) classic short-term memory task produces a pattern of reaction times (RTs) suggestive of serial exhaustive search through memory when subjects are given ample time to encode the memory set items and establish a rehearsal sequence. However, if the memory set items are presented quickly, RTs and accuracy reveal a parallel search process with a strong recency gradient (i.e., recently studied items are recognized more quickly and accurately). These phenomena have been well described for verbal material, but few studies have addressed the dynamics of rehearsal and search for purely visuospatial information and no studies have examined RTs following slow sequential presentation of the memory set. To address whether sequential rehearsal occurs for visuospatial information, we developed a novel visuospatial short-term memory task in which participants saw a sequence of colored dots (500 ms per dot) along a horizontal line and, after a brief delay (500 ms mask), gave a binary “old” / “new” response to a single test item. We compared four versions of the task, which varied in how similar the lures were to the items in the study sequence: lures were either a recombination of a previously viewed color and location (both old), a new color in an old location (new color), an old color in a new location (new location), or a new color in a new location (both new). Neither RTs nor accuracy revealed a pattern indicative of serial search in any of these conditions, suggesting a lack of sequential rehearsal. A recency gradient was observed for conditions in which color information was necessary (new color, both old), with negligible set size effects across the gradient. In contrast, the recency gradient was reduced for conditions where location information could serve as the sole dimension of evaluation (new location, both new), and set size effects were seen at each level of recency. These findings suggest that color information suffers from strong retroactive interference such that previous color representations are “overwritten,” whereas location representations are blurred with each presentation, as if the location information has been “averaged.”

Presenter: Dimitris Stergiedis

Title: Does providing a subtle reasoning hint remedy the conjunction fallacy?

Abstract: Humans are in general poor at making judgments that adhere to the logical principles of probability theory. One demonstration of this is termed the “conjunction fallacy”: judging a conjunction (A&B) as being more probable than its constituent (A). Systematic commitment of the conjunction fallacy has been shown in numerous studies on probability judgments. Different actions to remedy the fallacy have been suggested. According to the nested-sets hypothesis, when the nested-set structure of a problem becomes clear (i.e. the relation between categories and subcategories), then the conjunction fallacy is remedied. However, previous demonstrations of this remediation have provided very explicit task-related information and it can be questioned whether it is trivial that such information leads to more correct judgments. The primary aim of this study was to test the nested-sets hypothesis in two different formats of a probability judgment task, by more subtly hinting about the nested-set structure. Twenty-nine participants were randomly divided into two groups, one Probability condition and one Informed probability condition, where participants in the latter condition were provided with the hint. The second aim was to investigate whether the Informed probability condition was performed more slowly, potentially due to the time-cost of more elaborated judgments. The results show that a subtle hint about the nested-set structure was able to remedy the conjunction fallacy in a forced-choice probability judgment task but not statistically reliably in a probability estimation task. No response-time differences were observed between the conditions. The results support the nested-sets hypothesis and imply that even a subtle reasoning hint clarifying the relation between categories and subcategories might remedy one of the most robust probability judgment fallacies.

Friday April 28th at 3:30 in ILC S33, Moira Weigel of Yale University will be presenting a talk entitled “What is Political Correctness?” Weigel published a related article in the Guardian last November (https://www.theguardian.com/us-news/2016/nov/30/political-correctness-how-the-right-invented-phantom-enemy-donald-trump), and she’ll have more to say next Friday. The talk is being sponsored by the Departments of Political Science, Sociology, Languages, Literatures and Cultures, English, Communications, Linguistics, the Institute for Social Science Research, the College of Humanities and Fine Arts and the College of Social and Behavioral Science. A reception will follow in the Linguistics department.

Although this is not a CogSci talk, I thought it might be of general interest (Joe Pater)

Su-Lin Blodgett (CS) will be presenting in the next Computational Linguistics Community Meeting, in ILC N400 at 11:15 Friday April 28. Blodgett is working on developing models to identify dialectal variation on social media – see her EMNLP paper below.

Demographic Dialectal Variation in Social Media: A Case Study of African-American English. Su Lin Blodgett, Lisa Green, and Brendan O’Connor. Forthcoming, Proceedings of EMNLP 2016. [pdf] [site]

who: Tom Williams, Tufts University
when: 11:30, Wednesday, April 26
where: Computer Science Building Rm 150
food: Antonio’s pizza

Please note this MLFL is occurring on a different day and time:  Wednesday at 11:30 am.

“Genuine Helpers: Enabling Natural Language Capabilities for Interactive Robots”
ABSTRACT:  Natural language understanding and generation capabilities are crucial for natural human-like human robot interactions. This is especially true in domains such as eldercare, education, space, and search-and-rescue robotics, in which alternate interfaces or interaction techniques may be difficult for users to use due to cognitive or physical limitations. Approximately 40% of wheelchair users, for example, find it difficult or impossible to use a standard joystick, making natural language an attractive modality for interaction and control.
My research investigates how intelligent robots can communicate through natural language in realistic human-robot interaction scenarios, in which knowledge is uncertain, incomplete, and decentralized. To do so, I draw on techniques and concepts from artificial intelligence, psychology, linguistics, and philosophy, and engage in both algorithm development and empirical experimentation.
In my talk, I will present a set of cognitively inspired algorithms I have developed to allow robots to better identify the entities (e.g., objects, people, and locations) referenced in natural language by their human conversational partners, and to better infer those conversational partners’ intentions,  in uncertain and open worlds. I will then discuss how these algorithms have been implemented on a robotic wheelchair in order to significantly extend the state of the art of natural language enabled robot wheelchairs.
BIO: Tom Williams is a PhD candidate in the joint Computer Science and Cognitive Science program at Tufts University, and will be joining the faculty of Colorado School of Mines in the fall. Tom’s research focuses on enabling and understanding natural language based human-robot interaction, especially as applied to assistive and search-and-rescue robotics. He previously served as a visiting researcher at the Institute for Artificial Intelligence in Bremen, Germany, and has co-organized several international workshops on human-robot interaction.

Qiuli Ma (UMass) will give the next Cognitive Bag Lunch presentation at 12pm in Tobin 521B on Wednesday April 26th at noon (today!). The title and abstract follow. All are welcome!

Title: Testing recognition models with forced-choice testing

Abstract: We conducted two experiments to distinguish the double high-threshold model (2HT) and unequal variance signal detection model (UVSD) of recognition memory. During the experiments, participants studied a list of words. In the test phase, they first responded “old” or “new” to a single word in the recognition test. For a word that was incorrectly recognized, another word with the same response but was correctly recognized was paired with it to make a forced-choice trial. There were “new-new” (n-n) trials and “old-old” (o-o) trials. Participants were asked to pick the word that was actually studied from the two words. Data from the forced-choice testing were able to distinguish the 2HT and UVSD models. Modeling results showed that the UVSD model accounted for the forced-choice data better than 2HT model.

who: Hari Balasubramanian, UMass Amherst
when: noon, Thursday, April 20
where: Computer Science Building Rm 150
food: Antonio’s pizza

“Models Based on Longitudinal Healthcare Event Data”

Abstract: In this presentation, we will discuss a framework for analyzing data concerning healthcare events at the individual level. These events can be of various types – outpatient, emergency room, inpatient, lab, pharmaceutical etc., each corresponding to one or more diagnoses. Each event happens on a certain day (or a certain hour) and when such data is collected over a period of time, it creates an evolving point process unique to each patient. Such a point process provides information about the intensity and diversity of encounters – how frequent and how fragmented care is across multiple settings, an issue of particular concern for patients with multiple chronic conditions. In this presentation, we provide concrete examples of such datasets and the operational implications for clinicians. We will also try to seek the audience’s feedback on what machine learning techniques might be best suited to recognizing patterns in high-dimensional event sequences.

Bio: Dr. Hari Balasubramanian is Associate Professor of Industrial Engineering at the University of Massachusetts, Amherst. He received his doctoral degree at the Arizona State University in 2006. Dr. Balasubramanian spent two years as a Research Associate at Mayo Clinic in Rochester, Minnesota before joining the University of Massachusetts in 2008. His research interests are in operations research applied to healthcare. In 2013, Dr. Balasubramanian received a National Science Foundation CAREER award focused on healthcare delivery.