What: DS Seminar
Date: February 13, 2018
Time: 4:00 – 5:00 P.M.
Location: Computer Science Building, Room 151

Cameron Musco
The Power of Simple Algorithms: From Data Science to Biological Systems 

Abstract:  In recent years, very simple randomized methods, such as stochastic iteration, sampling, and hashing, have become dominant computational tools in large-scale machine learning and data science. In this talk, he will discuss his efforts to understand and harness the remarkable power of these methods.

In particular, he will describe his research on developing simple, but principled, sampling methods for learning, estimation, and optimization. He will present a new class of iterative sampling algorithms, which give state-of-the-art theoretical and empirical performance for regression problems, low-rank matrix approximation, and kernel methods. In many cases, the computational improvement offered by these algorithms is quite surprising. For example, our methods can be used to compute a near-optimal low-rank approximation to any positive semidefinite matrix in sublinear time.

In addition to their power in algorithm design, he will discuss his efforts to understand simple, randomized methods through a different lens: by studying how complex behavior emerges from low-level randomized interactions in biological systems. He will demonstrate how many of the same mathematical tools used to study algorithms in data science can be applied to these systems. As an example, he will highlight his research on noisy estimation and decision making in social insect colonies.

Bio:  Cameron Musco is a fifth year Ph.D. student (graduating spring 2018) in the Theory of Computation Group at MIT.
He is advised by Nancy Lynch and supported by an NSF Graduate Fellowship. He studies algorithms, focusing on applications in data science and machine learning. He often works on randomized methods and algorithms that adapt to streaming and distributed computation. He is also interested in understanding randomized computation and algorithmic robustness by studying computational processes in biological systems.

Before MIT, he studied Computer Science and Applied Mathematics at Yale University and worked as a software developer on the Data Team at Redfin.

Cognitive Brown Bag, 2/14/18, 12:00-1:15, Tobin 521B

Michele Fornaciai & Joonkoo Park (PBS)

Serial dependence in numerosity perception

Attractive serial dependence represents an adaptive change in the representation of sensory information, whereby current stimuli appear more similar to previous ones. Here, we characterize the behavioral and neural signatures of serial dependence in numerosity perception, demonstrating that the perceived numerosity of dot-array stimuli in different numerical ranges is biased by a preceding irrelevant stimulus (“inducer”) in an attractive way. Using electroencephalogram and a passive-viewing paradigm, we show that a neural signature of attractive serial dependence emerges even in the absence of an explicit task early in the visual stream, suggesting that serial dependence has a clear perceptual origin independently from a decision process. With a series of follow-up experiments, we further characterize serial dependence in visual number perception. First, we show that this effect has a weak spatial specificity and a relatively broad tuning for numerosity, and that it has a clear cortical origin (rather than subcortical). Second, we show that the attractive effect is strongly modulated by attention, suggesting the involvement of higher level modulatory influences. Our results collectively suggest that serial dependence results from a cortical neural computation starting from an early level of perceptual processing, possibly subserving perceptual stability and influencing downstream cognitive stages. However, these findings also suggest that the integration of past and present stimuli is in turn modulated by higher-level processes, and potentially amplified at later processing stages.

who: Bharath Hariharan, Cornell University
when: 11:45 A.M. 1:15 P.M., Thursday, February 8th
where: Computer Science Building Rm 150

Abstract: The performance of recognition systems has grown by leaps and bounds these last 5 years. However, modern recognition systems still require thousands of examples per class to train. Furthermore, expanding the capabilities of the system by introducing new visual concepts again requires collecting thousands of examples for the new concept. In contrast, humans are known to quickly learn new visual concepts from as few as 1 example, and indeed require very little labeled data to build their powerful visual systems from scratch. The requirement for large training sets also makes it infeasible to use current machine vision systems for rare or hard-to-annotate visual concepts or new imaging modalities.

I will talk about some of our work on reducing this need for large labeled training sets. I will describe novel loss functions for training convolutional network-based feature representations so that new concepts can be learned from a few examples, and ways of hallucinating additional examples for data-starved classes. I will also discuss our attempt to learn feature representations without any labeled data by leveraging motion-based grouping cues. I will end with a discussion of where we are and thoughts on the way forward.

Cognitive Brown Bag, 2/7/18, 12:00-1:15.

NOTE:  This week only, the talk will be in Bartlett 119

Ibrahim Dahlstrom-Hakki (Landmark College)

Teaching Students with Disabilities Online: Language-Based Challenges and Cognitive Access

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.

Machine Learning and Friends

who: Graham Neubig (CMU)
when: 11:45a – 1:15p, Feb 1st
where: Computer Science Building Rm 150
food: Antonios Pizza
generous sponsor:  ORACLE LABS 

What Can Neural Networks Teach us about Language?

Abstract:

Neural networks have led to large improvements in the accuracy of natural language processing systems. These have mainly been based on supervised learning: we create linguistic annotations for a large amount of training data, and train networks to faithfully reproduce these annotations. But what if we didn’t tell the neural net about explicitly, but instead *asked it what it thought* about language
without injecting our prior biases? Would the neural network be able
to learn from large amounts of data and confirm or discredit our
existing linguistic hypotheses? Would we be able to learn linguistic
information from lower-resourced languages where this information has not been annotated? In this talk, I will discuss methods for
unsupervised learning of linguistic information using neural networks
that attempt to answer these questions. I will also explain briefly
about automatic mini-batching, a computational method (implemented in the DyNet neural network toolkit), which greatly speeds large-scale
experiments with complicated network structures needed for this type
of unsupervised learning.

Cognitive Brown Bag, 1/31/18, 12:00-1:15, Tobin 521B.

Patrick Sadil   (PBS)

Pattern-completion of intra-item associations

Some neurocomputational theories of episodic memory have cast recollection as a pattern-completion process during which information is retrieved. In contrast to the pattern-matchingprocess of familiarity, recollection is thought to be most useful in tasks that require retrieval across episodic-like, arbitrary associations (e.g., retrieval of an item cued by a context). However, it is unclear whether pattern-completion occurs for the retrieval of purely visual, intra-item associations that exist at a lower level of representation in the brain’s processing hierarchy. In this experiment, we asked whether recollection, operationalized as the retrieval of previously studied material that is not provided as part of the cue, can occur for the features of single, everyday objects. Participants studied images of everyday objects. To manipulate the degree to which participants could encode arbitrary associations (e.g., the object with the experimental context) some objects were masked with continuous flash suppression. At test, participants were given a cued-recall test in which a small part of the studied object was presented and the task was to identify the whole object. Immediately after the cued-recall, participants engaged in a target detection go/no-go task where the inverse of the object part (i.e., the entire object save for the part used as a cue) might or might not emerge out of visual noise. The key question was whether the response time distribution was shifted earlier on go trials for objects that were studied under continuous flash suppression (as compared to a not-studied baseline) , in particular when the name of the object was not recalled immediately prior. Participants detected the non-recalled CFS-studied objects appearing from the noise more rapidly, and the speed-up corresponded to a shift in the response time distribution. This suggests that participants had a ‘head-start’ on the go trials, despite an inability to identify the cued object. This head-start suggests that a visual pattern-completion process occurred during cued-recall even when participants only encoded – and so could only retrieve – intra-item associations.

The Five College Cognitive Science Speaker Series and the UMass Initiative in Cognitive Science is pleased to announce that Sue Carey, Department of Psychology, Harvard, will speak at UMass on Friday, April 20 at 3:30pm in ILC S131. The title and abstract are below.

—————–

Do Non-Linguistic Creatures have a Fodorian (Logic-Like/Language-Like) Language of Thought?

The adult human conceptual repertoire is a unique phenomenon earth. Human adults build hierarchical representations on the fly, distinguishing “Molecules are made of tiny atoms” (True) from “Atoms are made of tiny molecules” (False). It is unknown whether non-linguistic creatures are capable of representing structured propositions in terms of hierarchical structures formulated over abstract variables, assigning truth values to those propositions, or are capable of abstract relational thought. How abstract knowledge and abstract combinatorial thought is acquired, over both evolutionary and ontogenetic time scales, is one of the outstanding scientific mysteries in the cognitive sciences, and has been debated in the philosophical literature at least since Descartes. Many philosophers, from Descartes through Davidson, have argued that abstract combinatorial thought is absent in creatures who lack natural language; others, such as Fodor, argue that such thought must be widely available to non-linguistic creatures, including human babies and animals at least throughout the vertebrates. A priori arguments will not get us far in settling this issue, which requires both theoretical analysis and empirical work. Theoretically, those who think there is a joint in nature between the kinds of representations that underlie perception and action, on the one hand, and abstract combinatorial thought, on the other, owe us an analysis of the essential differences between the representations and computations involved in each. Empirically, then, we must develop methods that could yield data that bear on the question of whether non-human animals or human infants have representations/computations on the abstract combinatorial thought side of the putative joint in nature. I will illustrate progress on both the theoretical and empirical fronts through two case studies: logical connectives and abstract relations.

Cognitive Brown Bag, 1/24/18, 12:00-1:15, Tobin 521B.

Speaker:  Chris Hammerly, UMass Linguistics

Susan Carey’s talk, scheduled for February 2nd, has been postponed. We are working on finding a new date.