STEM publications, for various reasons, generally do not place a premium on writing for readability, and young scholars/students struggle to understand the scholarly literature available to them. Unfortunately, few efforts have been made to help graduate students and other junior scholars understand and consume the essence of those scientific readings. This talk is based on the hypothesis and pilot-evidence that accessing multi-modal Open Data Resources (ODR) about a scholarly publication, including presentation videos, slides, tutorial, algorithm source code, or Wikipedia pages, in a collaborative framework will significantly enhance a student’s ability to understand the paper itself. To achieve this goal, I propose a novel learning/reading environment, ODR-based Collaborative PDF Reader (OCPR), that incorporates innovative text plus heterogeneous graph mining algorithms that can: 1) auto-characterize a student’s emerging information need while he/she reading a paper; 2) personalize or communitize students’ information needs based on the computational user profiles, and 3) enable students to readily access ODRs based on their information need and implicit/explicit feedback.
Based on the information need and various kinds of user feedback, the proposed algorithms will generate and select novel ranking features on the heterogeneous graph at a low cost for semi-supervised random walk and ODR recommendation. Experiment shows that the proposed system can effectively help graduate students and scholars better understand the complex publications in both cold start and context-rich environments, and the novel algorithms, e.g., personalized edge type usefulness estimation, can be generalized to other information recommendation/retrieval problems.

Even relatively simple animals exhibit a remarkable combination of flexibility and robustness to their behavior. One of the grand scientific challenges of this century is to understand how such adaptive behavior arises from the dynamical interaction of the organism’s nervous system, its body, and its environment. Although detailed reductionist analyses of the individual molecular, cellular and organismal components of biological systems have led to a remarkable wealth of data and insights throughout biology, a complementary synthetic approach that reintegrates these components into an understanding of whole systems has been lacking.
My research aims to address this challenge by constructing and analyzing empirically-grounded models of brain-body-environment systems. In this talk, I will introduce the nematode worm Caenorhabditis elegans as a convenient target for such integrated brain-body-environment modeling of a complete animal. I will describe my approach: using artificial evolution to explore the space of unknown electrophysiological parameters of the nervous system necessary to generate organism-like behavior. I will focus on the evolution and analysis of two of the worm’s behaviors: spatial orientation and locomotion, and their integration. I will show how this methodology allows us to begin to address key theoretical challenges in a situated, embodied, and dynamical understanding of cognition.

The project aims to develop and empirically test a new measurement model based on quantum probability theory, called the Hilbert space multi-dimensional model. The model provides a promising solution to the problems of the violations of joint distribution assumptions faced by complex data. With the striking advancement of modern data collection methods, complex and massive data sets are generated from various sources and contexts that are conceptually connected (e.g., big data). This promises to provide a better understanding of complex social and behavioral phenomena, but also presents unprecedented challenges for the integration and interpretation of the data. When large data sets are collected from different contexts, often they can be summarized by contingency tables. Suppose there are K tables (T1,…,Tk,…TK), each collected in a different context k. Also suppose that each table Tk is a joint frequency table based on a subset of p variables (Y1, …, Yp). For example, the research could involve four variables (Y1, Y2, Y3, Y4), but each table might include only two of the four, so that table T1 might be a 2-way frequency table composed of two variables (Y1, Y2), table T2 might be another 2-way table composed of another two variables (Y1, Y3), and so on. A critical problem arises: How to integrate and synthesize these K tables into a compressed, coherent, and interpretable representation?
Currently, a common solution is to try to construct a p-way joint probability distribution to reproduce the frequency data observed in the K tables. Often Bayesian causal networks are then used to reduce the number of estimated parameters by imposing conditional independence assumptions. Unfortunately, however, in many cases, no such p-way joint distribution exists that can reproduce the observed tables. This occurs because the data tables violate consistency constraints required by classical (Kolmogorov) probability theory that Bayes nets are built upon. Research supported by prior NSF grants has accumulated strong evidence for the violations of the often-assumed classical joint probability idea on complex, contextualized data. The Hilbert space model, as we proposed here, is based on quantum probability theory. It provides a promising solution to the problems of the violations of joint distribution assumptions by constructing a single finite state vector that lies within a low dimensional Hilbert space, and by forming a set of non-commuting measurement operators that represent the p measurements. In this way, we achieve a compressed, coherent, and interpretable representation of the p variables that form the complex collection of K tables even when no p-way joint distribution
exists.

Every day billions of images are uploaded to the Internet. Together they provide many high-resolution pictures of the world, from panoramic views of natural landscapes to detailed views of what someone had for dinner. This imagery has the potential to drive discoveries in a wide variety of disciplines, from environmental monitoring to cultural anthropology. Significant research progress has been made in automatically extracting information from such imagery. One of the key remaining challenges is that we often don’t know where an image was captured and usually know very little about other geometric properties of the camera, such as orientation and focal length. In other words, most images are not geocalibrated. This talk provides an overview of my work on using novel cues, including partly cloudy days, rainbows, and human faces, to geocalibrate Internet imagery and video

There is a wide gap in the field of machine learning, between the language that is used to describe and share successful ideas, and the code that is used to execute them. While the language of machine learning allows us to reuse a set of concepts, the same cannot be said of the code. For example, if we wanted to query an image processing model by using an inference technique originally written for a speech recognition model, we’d likely have to write an inference method from scratch even if the models shared structural similarities. Instead, we would like small conceptual changes to lead to proportionally small code changes, which is the basis of programming more modularly.

A first step towards modularity is to decouple
– the code describing the model from
– the code describing the inference method.
This type of modularity is a direct result of programming in the “generative story” style of probabilistic programs. A second step towards modularity is to compose bigger models from smaller (working) models, and to similarly compose inference methods.
This talk will consider modularity in machine learning programs by surveying the generative story style of programming and examining the tools for model and inference composition provided by Hakaru, a probabilistic programming system being developed at Indiana.

Exponential random graph models (ERGMs) have become a standard statistical tool for modeling social networks. In particular, ERGMs provide great flexibility to account for both covariates effects on tie formation and endogenous network formation processes (e.g., reciprocity and transitivity). However, due to the reliance on Monte Carlo Markov Chains, it is difficult to fit ERGMs on large networks (e.g., networks composed of hundreds of nodes and edges). This paper describes a series of (existent and new) methods for estimating ERGMs on large networks and compares their advantages and disadvantages. Selected methods are illustrated through analyzing school friendship networks, etc.

Probabilistic inference procedures are usually coded painstakingly from scratch, for each target model and each inference algorithm. In this talk, I will show how inference procedures can be posed as program transformations which transform one probabilistic model into another probabilistic model. These transformations allow us to generate programs which express exact and approximate inference, and allow us to compose multiple inference procedures over a single model. The resulting inference procedure runs in
time comparable to a handwritten procedure.

Formal notations like algebraic notations are powerful tools of thinking that can vastly extend our cognitive capabilities. Yet, many students don’t get much out of using them except frustration. While
many of our reasoning tools have changed dramatically with the availability of computers and digital user interfaces, we still mostly write equations using pen and paper.
A few years ago, our team set out to design and implement a digital math notation, inspired by modern user interface design and guided by cognitive research on how people use their visual-motor systems to work with formal notations. Since then, we have built a web-based dynamic algebra notation system we call Graspable Math. With it, users can solve a subset of algebra problems faster and more accurately than on paper, while still working through all transformation steps themselves.
In our talk, we will briefly discuss the research and motivation behind building Graspable Math, demonstrate the system, and share some details about its architecture and our software development strategies. We’ll then discuss our plans for the future and some of the challenges that lie ahead.

(Check out our system at http://graspablemath.com)

Search in an environment with an uncertain distribution of resources involves a trade-off between exploitation of past discoveries and further exploration. This extends to information foraging, where a knowledge-seeker shifts between reading in depth and studying new domains. To study this cognitive process, we examine the reading choices made by one of the most celebrated scientists of the modern era: Charles Darwin. From the full-text of books listed in his chronologically-organized reading journals, we generate topic models to quantify his local (text-to-text) and global (text-to-past) reading decisions using Kullback-Liebler Divergence, a cognitively-validated, information-theoretic measure of relative surprise. Rather than a pattern of surprise-minimization, corresponding to a pure exploitation strategy, Darwin’s behavior shifts from early exploitation to later exploration, seeking unusually high levels of cognitive surprise relative to previous eras. These shifts, detected by an unsupervised Bayesian model, correlate with major intellectual epochs of his career as identified both by traditional, qualitative scholarship and Darwin’s own self-commentary. In addition to quantifying Darwin’s individual-level foraging, our methods allow us to compare his consumption of texts with their publication order. We find Darwin’s consumption more exploratory than the culture’s production, suggesting that underneath gradual societal changes are the explorations of individual synthesis and discovery. Our quantitative methods advance the study of cognitive search through a framework for testing interactions between individual and collective behavior and between short- and long-term consumption choices. This novel application of
topic modeling to characterize individual reading complements widespread studies of collective scientific behavior.

Paper preprint: http://arxiv.org/abs/1509.07175

Emerging technologies have revealed many features of the behavior and structure of cortical microcircuits. We now know that neuronal firing rates are roughly log-normal and that synaptic connectivity is highly non-random, containing over represented motifs, hubs, and heavy-tailed distributions in both synaptic efficacy and degree. Recent self-organizing neural models have begun to use known homeostatic mechanisms in combination with other forms of neural plasticity to explain these phenomena. However, while many of these models take into account homeostasis few put a great deal of consideration into how the set-points of said homeostasis are reached. They–therefore–largely ignore developmental differentiation as a key aspect of self-organization in the brain. The following talk presents a novel neural circuit model wherein the individual set-points of each neuron’s homeostatic mechanisms are allowed to self-organize alongside the network’s excitatory and inhibitory synaptic growth. Along those same lines, the network is initialized with no pre-defined synaptic connectivity and must grow and then prune its synaptic connections. From there, the network self-organizes a modular/hierarchical structure, rich-club organization, and broadly emulates other qualities of living circuits such as their degree, versatility, synaptic efficacy, and 3-motif distributions. The network also shows an ability to optimally conform around the structure of its inputs, scoring high on measures of pattern separation or generalization depending on which task is favored by the inputs. The results demonstrate that the self-organization of homeostatic set points in a cortical model can reproduce a plethora of highly non-random network
features known to exist in living neural circuits.

Understanding how an artificial agent may represent, acquire, update, and use large amounts of knowledge has long been an important research challenge in artificial intelligence. This talk explores the predictive approach to knowledge. Predictive knowledge can be maintained without human intervention, and thus its acquisition can potentially scale with available data and computing resources. Unfortunately, technical challenges related to numerical instability, divergence under off-policy sampling, and computational complexity have limited the applicability and scalability of predictive knowledge acquisition in practice.
This talk describes a new approach to representing and acquiring predictive knowledge on a robot. The key idea is that value functions, from reinforcement learning, can be used to represent policy-contingent declarative and goal-oriented predictive knowledge. This talk explores the practicality of making and updating many predictions in parallel, while the agent interacts with the world. I will demonstrate the applicability and scalability of our approach with a demonstration of the psychological phenomenon of nexting, making and updating thousands of predictions from hundreds of thousands of multi-dimensional data samples, in realtime and on a robot—beyond the scalability of related predictive approaches.

In this talk, we highlight the interplay between a dynamic process and the structure of the network on which it is defined. We start by examining the impact of different random walks on the quality-measure of network clusters and node centrality. We introduce an umbrella framework for defining and characterizing an ensemble of dynamic processes as linear operators. We show that the traditional Laplacian framework for diffusion and random walks is a special case of this framework. Further generalizations will allow us to model epidemic and information diffusion over a networks.
Based on this generalized Laplacian framework, we will demonstrate how linear transformations of graphs can represent the flow of different dynamic processes on networks. We will show some empirical examples of how such transformations can be applied in real world problems where additional data is available along side the network structure. In the case of multiple graphs, the transformations lead to more principled composition of multi-layered networks.

With growing concerns about environment, economy, sustainability, and security, more and more renewable energy resources are expected to be integrated into future electrical grids. However, large-scale integration of intermittent renewables such as wind and solar creates noticeable imbalance between supply and demand, hence jeopardizing grid reliability. To combat the variability of renewable generation, energy storage (e.g., batteries, pumped hydro storage, and compressed air energy storage) and flexible loads (e.g., heating, ventilation, and air conditioning) are suggested to be applied in many grid-wideservices.

In this talk, I will present my work on intelligent control of storage and flexible loads in renewable-integrated electrical grids. First, I introduce an aggregator-storage system that provides service of power balancing to a grid. Both static and dynamic storage (e.g., storage inside electric vehicles) are considered with a wide range of storage characteristics being explicitly modeled. Second, for a substation to maintain phase balance, I suggest intelligent control of storage charging and discharging to balance energy flows. Third, with the inclusion of flexible loads in energy management, I propose the joint optimization of supply, demand, and storage for power balancing in a grid. To improve long-term system performance (e.g., reliability, welfare, and cost effectiveness), in each case I offer efficient centralized algorithms with strong theoretical performance guarantee, and distributed implementation with limited requirement of information exchange.

The emergence of social media has had a significant impact on how people communicate and socialize. Teens use social media to make and maintain social connections with friends and build their reputation. Research has suggested that teens are more active and engaged than adults on social media. Most of such observations, however, have been made through the analysis of limited ethnographic or cross-sectional data. This paper shows the possibility of detecting age information in user profiles by using a combination of textual and facial recognition methods and presents a comparative study of 27K teens and adults in Instagram. We examined how and why the age difference in the behaviors of users in Instagram might have occurred through the lenses of social cognition, developmental psychology, and human-computer interaction. We proposed two hypotheses — teens as digital natives and the need for social interactions — as the theoretical framework for understanding the factors that help explain the behavioral differences. We demonstrate the application of our novel method that shows clear trends of age differences as well as substantiates previous insights in social media. Our computational analysis identified the following novel findings: (1) teens post fewer photos than adults; (2) teens remove more photos based on the number of Likes the photos received; and (3) teens have less diverse photo content. Our analysis was also able to confirm prior ethnographic accounts that teens are more engaged in Liking and commenting, and express their emotions and social interests more than adults. We discussed theoretical and practical interpretations and implications as well as future research directions from the results.

I’ll discuss my work with musical accompaniment systems both giving live and video demonstrations as well as presenting the underlying modeling.

Can robots understand human activities based on their visual perception? How can we make them do so? This talk discusses Computer Vision algorithms necessary to provide activity-level situation awareness to robots. The activities a robot must recognize not only include simple actions by users in front of it, but also include interactions directly involving the robot such as ‘a person attacking the robot’.  The objective is to reliably handle such videos obtained during human/robot activities (which often display significant ego-motion) and associate semantic annotations. The videos taken from a robot’s own perspective are called first-person (or egocentric) videos and we review methods necessary for them. The talk also overviews ongoing work on robot learning of ‘actionable’ activity representations.

Sentiment Analysis, also known as opinion mining, is the task of extracting opinions/emotions from user generated text, including blogs, conversations, reviews, etc. In this talk, we investigate feature selection methods for sentiment analysis. In particular we address two questions: 1) Feature selection as a mechanism to cut down the high dimensional feature vectors by identifying salient features, are normally defined for binary classification; we will investigate what is the best method to extend them to a multi-class setting. 2) One issue that is pervasive but usually overlooked in sentiment analysis research is imbalanced data; we will investigate whether the selected features are representative of the minority class, and whether we could mitigate the effect of skewing by making the features more general.

In the near future, every house is predicted to have multiple robots: One welcomes you home, another washes your dishes, and a third folds laundry. Researchers and funders alike expect that humans and robots will be able to “symbiotically coexist” and collaborate; however, researchers have yet to study whether and how the differences we commonly see between individual and intergroup interactions among humans will affect human-robot interaction.
I will discuss my research, which begins to examine how interactions with groups of robots in an intergroup context may differ from one-on-one interactions. I will finish by discussing what future directions this research should include. Overall, the goal is to make human-robot collaboration more effective, efficient, and pleasant.

Building human-in-the-loop intelligent systems that effectively utilize expert feedback to learn robust decision models has been a long cherished goal of Artificial Intelligence. However, modern decision systems are overly reliant on ideal training data and
consequently, any imperfections in the data will be reflected in the final model. It is vital that many real world decision systems make use of the vast amount of expert knowledge gained in fields such as
robotics, navigation and personal healthcare assistants.
First, in this talk, I will outline our algorithm for exploiting domain knowledge in sequential decision making systems. While the acquisition of domain knowledge is important, many domain experts are not machine learning experts and cannot define the most useful feedback. Ideal algorithms can communicate to the expert where they require assistance. In the second half of the talk, I present a method that, by actively soliciting expert feedback, trades off between the performance of the system and the effort required by the expert. We show empirically the contribution of more expressive advice over traditional learning approaches.

Estimation of class proportions is one of the most fundamental tasks in machine learning yet this research problem is surprisingly open in many applications. In this presentation I will discuss supervised, semi-supervised and unsupervised approaches to these estimation problems, with a particular focus on semi-supervised techniques in which only positive and unlabeled data are available. There, we are interested in non-parametric techniques for class prior estimation. We recently formulated this problem as an estimation of mixing proportions in two-component mixture models. We then showed that estimation of mixing proportions is generally ill-defined and proposed a canonical form to obtain identifiability while maintaining flexibility to model any distribution. We used insights from this theory to elucidate the optimization surface of the class priors and proposed an algorithm for estimating them. The efficacy of the approach was evaluated on univariate and multivariate data. At the end, I hope to show some fun results obtained on estimation problems in bioinformatics where we do not know the truth. This is a joint work with Shantanu Jain, Martha White, and Michael Trosset.

We develop parametric and nonparametric algorithms to estimate the mixing proportion of a two component mixture given a sample from the
mixture and a sample from one of the components. This problem occurs in many domains, and has implications towards binary classification in the positive and unlabeled data setting. In the parametric setting, the components are assumed to have a skew normal distribution. In general, the problem is ill-defined, in the sense that the mixing proportion is not identifiable. We develop conditions that lead to identifiability. To handle multi-dimensional data, we give a transform that reduces the data to one-dimension and preserves the mixing proportion as well.

Principal components analysis (PCA) is a classical dimension reduction method that involves the projection of the data onto the subspace spanned by the leading eigenvectors of the covariance matrix. This projection can be used either for exploratory purposes or as an input for further analysis, e.g. regression. If the data have billions of entries or more, the computational and storage requirements for saving and manipulating the design matrix in fast memory is prohibitive. Recently, the Nyström and column-sampling methods have appeared in the numerical linear algebra community for the randomized approximation of the singular value decomposition of large matrices. However, their utility for statistical applications remains unclear. We compare these approximations theoretically by bounding the distance between the induced subspaces and the desired, but computationally infeasible, PCA subspace. Additionally we show empirically, through simulations and a real data example, the trade-off of approximation accuracy and computational complexity.

Social robotics is one of the main areas of robotics development in Japan, particularly in the domain of personal robots for everyday use. This talk will first report on the current state of the art in Japanese robotics, including commercialized platforms like Aldebaran’s Pepper humanoid, Palro, and the seal-like robot Paro. We will then describe our lab’s work on designing a robot prototype “Mugbot” as an example of how a social robot can be developed through interaction with the community, and how the use of a robot shapes community in turn. Mugbot has been used by children in educational contexts, and has also been presented at several Maker Faires as a DIY robotics platform.
Bio: Seita Koike is a Professor of Informatics at Tokyo City University. He founded and directs the Information Design Laboratory. The lab’s research focuses on information design, social robotics, education, and human-robot interaction. After working for many years with NEC’s personal robot Papero, Dr. Koike developed his own prototype social robot, Mugbot, which won four Blue Ribbon Maker Faire awards in 2012. Dr. Koike got his PhD in Industrial Design from Chiba University in 1986.

The environmental monitoring field often suffers from lack of significant and exhaustive input data. On the other hand, the quantity of public content generated by users or by sensors available on the Web nowadays is reaching unprecedented volumes. This massive collection of data contains an enormous amount of latent knowledge, which can be used to improve environmental models. In particular, computer vision techniques can be applied to visual data, in order to extract relevant environmental information. For example, estimating snow cover in mountainous regions, that is, the spatial extent of the earth surface covered by snow, is an important challenge for efficient environmental monitoring and water management. Publicly available visual content, in the form of user generated photographs and image feeds from outdoor webcams, can be leveraged as additional measurement sources, complementing existing ground, satellite and airborne sensor data. Our SnowWatch platform implements two content acquisition and processing pipelines that are tailored to such sources, addressing the specific challenges posed by each of them, e.g., identifying the mountain peaks, filtering out images taken in bad weather conditions, handling varying illumination conditions, and classifying a pixel as snow or non-snow area. The final outcome is summarized in a snow cover index, which indicates for a specific mountain and day of the year, the fraction of visible area covered by snow, possibly at different elevations. Feeding snow cover indexes to real environmental models reveals environmental consistency and utility of the produced data.

Isomap, Locally Linear Embedding (LLE), and Laplacian Eigenmaps were originally promoted as techniques for nonlinear dimension reduction. Efforts to develop theories of manifold learning have assumed that data lie on low-dimensional manifolds and that the goal of manifold learning is to identify these manifolds. I will argue that this goal is conceptually distinct from the goal of nonlinear dimension reduction. For example, Hessian eigenmaps can (in theory) recover parametrizations that Isomap cannot, but Isomap can construct low-dimensional Euclidean representations of Riemannian manifolds that elude Hessian eigenmaps.

Balancing between computational efficiency and sample efficiency is an important goal in reinforcement learning. Temporal difference learning algorithms stochastically update the value function, with a linear time complexity in the number of features, whereas least-squares temporal difference algorithms are sample efficient but can be quadratic in the number of features. In this talk, I will discuss recent progress towards the goal of better balancing computation and sample efficiency with a new class of algorithms that use incremental matrix approximations.

Statistical Relational Learning (SRL) Models combine the powerful formalisms of probability theory and first-order logic to handleuncertainty in large, complex problems. While they provide a very effective representation paradigm due to their succinctness and parameter sharing, efficient learning is a significant problem in these models. First, I will discuss state-of-the-art learning method
based on boosting that is representation independent. Our results demonstrate that learning multiple weak models can lead to a dramatic improvement in accuracy and efficiency.

One of the key attractive properties of SRL models is that they use a rich representation for modeling the domain that potentially allows for seam-less human interaction. However, in current SRL research, the human is restricted to either being a mere labeler or being an oracle who provides the entire model. I will present our recent work that allows for more reasonable human interaction where the human input is taken as “advice” and the learning algorithm combines this advice with data. Finally, I will discuss our work on employing SRL models for achieving transfer across seemingly unrelated domains allowing for more efficient learning in data-scarce domains.