Hans Kamp  (Stuttgart/University of Texas at Austin) will present his forthcoming work on definites in a special colloquium on Friday, April 21, from 3:00 to 6:00 pm in N400. The colloquium is part of Daniel Altshuler’s graduate semantics seminar.

Please join us for a special talk by Randy Gallistel (Rutgers U.), 12-1 in the Cape Cod Lounge in the Student Union. Sponsored by the UMass Initiative in Cognitive Science and by the Five College Cognitive Science speaker series.

Title: It’s the neuron! How the brain really works.

Abstract: It is generally assumed that the brain’s computational capacities derive mostly from the structure of neural circuits—how it is wired—and from process(es) that rewire circuits in response to experience. The computationally relevant properties ascribed to the neuron itself have not changed in more than a century: It is a leaky integrator with a threshold on its output (Sherrington, 1906). The concepts at the core of molecular biology were undreamed of in Sherrington’s philosophy. They have transformed biological thinking in the last half century. But they play little role in theorizing about how nervous tissue computes. The possibility that the neuron is a full-blown computing machine in its own right, able to store acquired information and to perform complex computations on it, has barely been bruited. I urge us to consider it.

My reasons are: 1) The hypothesis that acquired information is stored in altered synapses is a conceptual dead end. In more than a century, no one has explained even in principle how altered synapses can carry information forward in time in a computationally accessible form. 2) It is easy to suggest several different models for how molecules known to exist inside cells can carry acquired information in a computationally accessible form. 3) The logic gates out of which all computation may be built are known to be implemented at the molecular level inside cells. Implementing memory and computation at the molecular level increases the speed (operations/s), energy efficiency (operations/J) and spatial efficiency (bits/m3) of computation and memory by many orders of magnitude. 5) Recent experimental findings strongly suggest that (at least some) memory resides inside the neuron.

Jessica Coon of McGill University will be presenting “Building verbs in Chuj: Consequences for the nature of roots” in the Linguistics colloquium series Friday April 14th at 3:30, in ILC N400. All are welcome!

Abstract: The suffix -w in Chuj (Mayan) is found in two contexts: (i) attached to transitive roots to form what have been called “incorporation antipassives” and (ii) attached to nominal and positional roots to form unergatives. In both contexts, the result is an intransitive verb with a single, agentive external argument. In this talk I provide a unified analysis of these constructions in which -w is a v/Voice head that attaches to a root and introduces an external argument, but does not assign ergative case. Intransitivity is indirectly ensured through the limited availability of licensing heads. This has important implications for the status of certain antipassives. In Chuj, I argue that the incorporation antipassive formed with -w does not convert a transitive verb into an intransitive verb, but rather, both transitive and “antipassive” stems are formed directly from the root.

This detailed look at Chuj verbal morphology sets the scene for a broader question: when it comes to verb-stem formation, what information is contributed by the root, and what is contributed by the functional heads? I argue first that roots in Chuj are not acategorical, but must be grouped into categories based on their stem-forming possibilities. Root category does not map directly to surface category, but does determine which functional heads are possible. Second, I show that while licensing, agreement, and the introduction of the external argument are all governed by higher functional heads, the presence or absence of an internal argument is dictated by the root. Specifically, transitive roots in Chuj always combine with an internal argument, whether it be a full DP, a bare pseudo-incorporated NP, or an implicit object. The internal argument cannot be fully omitted or suppressed, regardless of higher functional material.

who: Sainbayar Sukhbaatar, New York University
when: noon, Thursday, April 13
where: Computer Science Building Rm 150
food: Antonio’s pizza
generous sponsor: ORACLE LABS

Intrinsic Motivation And Automatic Curricula Via Asymmetric Self-Play

Abstract: We describe a simple scheme that allows an agent to explore its environment in an unsupervised manner. Our scheme pits two versions of the same agent, Alice and Bob, against one another. Alice proposes a task for Bob to complete; and then Bob attempts to complete the task. In this work we will focus on (nearly) reversible environments, or environments that can be reset, and Alice will “propose” the task by running a set of actions and then Bob must partially undo, or repeat them, respectively. Via an appropriate reward structure, Alice and Bob automatically generate a curriculum of exploration, enabling unsupervised training of the agent. When deployed on an RL task within the environment, this unsupervised training reduces the number of episodes needed to learn.

Bio: I am a PhD student working with Rob Fergus and Yann Lecun at Department of Computer Science in New York University. My research interests are deep learning, neuroscience and reinforcement learning. http://cims.nyu.edu/~sainbar/

Marc Howard (Boston U.) will give the next Cognitive Bag Lunch presentation at 12pm in Tobin 521B Weds. April 12 at noon. Title and abstract below – all are welcome!

Title: Retrieval of temporal context in the brain is associated with high confidence recognition responses

Abstract: Over the last several decades, models of human episodic memory tasks have bifurcated into models of recall and models of item recognition. Many models of recall have relied heavily on recovery of a gradually-changing state of temporal context. This putative mechanism is sufficient to account for the contiguity effect in recall tasks. In contrast, process models of recognition have focused on accessing traces and matching features. We review recent results from neurobiological data demonstrating that the ensemble response in the brain changes across a wide range of time scales from seconds up to at least several days. We present unpublished results from a study using single unit recordings from an item recognition experiment. Human epilepsy patients studied a list of pictures. The single-unit response changed gradually across stimulus presentations, as predicted by models of temporal context. When an old item was presented as a probe and that probe received a highest confidence response, the neural ensemble “jumped back in time”, recovering the prior state of temporal context. When an old probe did not receive a highest confidence response, the effect was not observed, and in fact showed a weak anti-contiguity effect. We review several unpublished behavioral results and discuss a research strategy for constructing unified models of episodic memory performance that include both recall and recognition results.

The location for the Gallistel talk has been changed to the Cape Cod Lounge of the Student Union Building. Other details are below.

Charles Randy Gallistel, Distinguished Professor Emeritus of Psychology at Rutgers University, will be the 5 Colleges Cognitive Science Speaker this year, in a talk co-sponsored by the Initiative in Cognitive Science. He will present “It’s the neuron! How the brain really works” in the Cape Cod Lounge of the Student Union Building from noon to 1 pm on Wednesday April 19. An abstract and a poster follow.

Abstract. It is generally assumed that the brain’s computational capacities derive mostly from the structure of neural circuits—how it is wired—and from process(es) that rewire circuits in response to experience. The computationally relevant properties ascribed to the neuron itself have not changed in more than a century: It is a leaky integrator with a threshold on its output (Sherrington, 1906). The concepts at the core of molecular biology were undreamed of in Sherrington’s philosophy. They have transformed biological thinking in the last half century. But they play little role in theorizing about how nervous tissue computes. The possibility that the neuron is a full-blown computing machine in its own right, able to store acquired information and to perform complex computations on it, has barely been bruited. I urge us to consider it.

My reasons are: 1) The hypothesis that acquired information is stored in altered synapses is a conceptual dead end. In more than a century, no one has explained even in principle how altered synapses can carry information forward in time in a computationally accessible form. 2) It is easy to suggest several different models for how molecules known to exist inside cells can carry acquired information in a computationally accessible form. 3) The logic gates out of which all computation may be built are known to be implemented at the molecular level inside cells. Implementing memory and computation at the molecular level increases the speed (operations/s), energy efficiency (operations/J) and spatial efficiency (bits/m3) of computation and memory by many orders of magnitude. 5) Recent experimental findings strongly suggest that (at least some) memory resides inside the neuron.

Alexandra Jesse (UMass) will give the next Cognitive Bag Lunch presentation at 12pm in Tobin 521B on Wednesday April 5th at noon. The title and abstract follow. All are welcome!

Learning about speaker idiosyncrasies in visual speech

Seeing a speaker typically improves speech perception, especially in adverse conditions. Audiovisual speech is more robustly recognized than auditory speech, since visual speech assists recognition by contributing information that is redundant and complementary to the information obtained from auditory speech. The realization of phonemes varies, however, across speakers, and listeners are sensitive to this variation in both auditory and visual speech during speech recognition. But listeners are also sensitive to consistency in articulation within a speaker. When an idiosyncratic articulation renders a sound ambiguous, listeners use available disambiguating information, such as lexical knowledge or visual speech information, to adjust the boundaries of their auditory phonetic categories to incorporate the speech sound into the intended category. This facilitates future recognition of the sound. For visual speech to best aid recognition, listeners likewise have to flexibly adjust their visual phonetic categories to speakers. In this talk, I will present work showing how lexical knowledge and auditory speech information can both assist the retuning of visual phonetic categories to speakers, but that at least the latter type of retuning seems to rely on attentional resources. Furthermore, listeners rapidly form identity representations of unfamiliar speakers’ facial motion signatures, which subserve talker recognition but may also aid speech perception.

Talk postponed until fall

Mary Farbood of NYU will present a talk in room 272 of the Fine Arts Center’s music wing at 2:30 of Friday March 31. A title and abstract follow.

Title: The temporal dynamics of music versus speech processing

Abstract: 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.

Charles Randy Gallistel, Distinguished Professor Emeritus of Psychology at Rutgers University, will be the 5 Colleges Cognitive Science Speaker this year, in a talk co-sponsored by the Initiative in Cognitive Science. He will present “It’s the neuron! How the brain really works” in Integrative Learning Center room N400 from noon to 1 pm on Wednesday April 19. An abstract and a poster follow.

Abstract. It is generally assumed that the brain’s computational capacities derive mostly from the structure of neural circuits—how it is wired—and from process(es) that rewire circuits in response to experience. The computationally relevant properties ascribed to the neuron itself have not changed in more than a century: It is a leaky integrator with a threshold on its output (Sherrington, 1906). The concepts at the core of molecular biology were undreamed of in Sherrington’s philosophy. They have transformed biological thinking in the last half century. But they play little role in theorizing about how nervous tissue computes. The possibility that the neuron is a full-blown computing machine in its own right, able to store acquired information and to perform complex computations on it, has barely been bruited. I urge us to consider it.

My reasons are: 1) The hypothesis that acquired information is stored in altered synapses is a conceptual dead end. In more than a century, no one has explained even in principle how altered synapses can carry information forward in time in a computationally accessible form. 2) It is easy to suggest several different models for how molecules known to exist inside cells can carry acquired information in a computationally accessible form. 3) The logic gates out of which all computation may be built are known to be implemented at the molecular level inside cells. Implementing memory and computation at the molecular level increases the speed (operations/s), energy efficiency (operations/J) and spatial efficiency (bits/m³) of computation and memory by many orders of magnitude. 5) Recent experimental findings strongly suggest that (at least some) memory resides inside the neuron.

Junha Chang (UMass) will present in the Cognitive Bag Lunch Wednesday, March 29 at 12pm in Tobin 521B. All are welcome! Title and abstract follow.

Title: Search guidance can be adjusted by experience with search discriminability

Abstract: Several recent studies show that previous experience can influence search strategy in a way that improves search performance. The purpose of the present study is to investigate how the experience of difficult color discriminability affects search strategies. Two participants groups either experienced difficult color discriminability in a half of the trials (i.e., hard-discrimination group) or experienced easy search in all trials (i.e., easy-discrimination group) in a dual-target search task. Participants were required to respond to the presence of a target (colored T) among distractors (colored pseudo-L). Eye movements were recorded to understand which feature information is used to guide attention, and response times and error rates were measured to compare search efficiency between the two participant groups. The hard-discrimination group fixated more distractors with target-dissimilar colors than the easy-discrimination group, suggesting the hard-discrimination group used shape information to guide search more than the easy-discrimination group. However, the error rates and response times were not significantly different between groups. The results demonstrate that the experience of difficult color discriminability discourages participants from guiding attention by color, and encourages them to use shape information.