Michael C. Mozer - US grants

The connectionist paradigm provides the initial assumptions which form the basis of this work on characterizing mechanisms of the mind in terms of large networks of autonomous processing elements. Specific aspects of this work include models of human cognition, connectionist AI, and connectionist lerning algorithms in an interleaved fashion. This work focuses specifically on attention which will be studied in perception and also in higher levels of cognition.

The interdisciplinary field of "connectionism", or "neural networks," explores computer systems that are loosely modeled on the human brain. These systems, composed of large numbers of simple neuron-like computing elements, exhibit a variety of intelligent behaviors, including the ability to recognize visual images (e.g., handwriting recognition), store and remember facts by association, control complex mechanical systems (e.g., robot arms), understand natural language, and learn from experience. The field has experienced rapid growth in recent years and is now enjoying an explosion of popularity, due in large part to the rapid rate of advances and practical applications of recent theoretical developments. Connectionism provides a possible approach to solving complex problems in artificial intelligence and pattern recognition for which no other approaches appear viable. Connectionist models are being used successfully in the real world and have opened new possibilities for explaining aspects of human cognition and perception. Research in the field requires remarkably broad and deep training in a number of areas--neurobiology, cognitive science, theoretical foundations, computational methods, engineering applications, and hardware implementation. Some students are fortunate enough to receive training from leaders in one or two areas, but few if any have access to the full breadth of high-quality training required. For this reason, Summer Schools have been organized to train top young researchers and to encourage interdisciplinary research collaboration. These Summer Schools provide an intense and stimulating opportunity for graduate students to become immersed in the latest research in this fast-changing field. The success of the previous Summer Schools has been widely acknowledged; they have had an extremely important catalytic effect on the field. Many of the leading young researchers in the field today attended a previous Summer School, and their research directions and contributions were influenced by the experience. The next Summer School will be held in Boulder, Colorado from June 21 through July 3, 1993. The faculty of the 1993 Summer School will consist of approximately twenty leading researchers from the US and Canada. Sixty-five graduate students will be selected by a competitive admissions process for the 12-day intensive program. This funding will allow graduate students to be selected on the basis of merit and potential for contributing to the field.

9422202 Mozer Applicants to the NSF 1994 Minority Graduate Fellowship competition who were awarded "Honorable Mention" status and who enrolled in a computer science or computer engineering graduate program at a U.S. university were eligible to apply to the CISE Directorate for this special award. The purpose of the award is to assist the student in both research and educational activities related to his/her graduate education. The award is made on behalf of the student to the institution with the student's advisor designated as principal investigator. ***

This project aims to futher our understanding of the neural representations that underlie working memory. Working memory refers to the active maintenance of information in the service of complex cognitive tasks such as problem solving and planning. The team of investigator with study how the unique demands placed on the working memory system shape its representations during learning and development, and how these representations affect the use of working memory by the cognitive system as a whole. Their primary source of insight into this process will come from a computational analysis, which will be used to integrate and explore relevant findings from neurobiology and developmental and adult cognition. The primary hypothesis is that to maintain information in an active state over delays and in the face of interference, working memory representations should be discrete. A discrete representation admits to only a finite set of possible states, rather than representing continuous states. For example, the integers from 1 to 100 form a discrete set, in contrast to the real numbers in this range. Discreteness imparts a measure of robustness to the representation because small amounts of "noise" can be overcome by interpreting an observed state as the nearest discrete stat. From the central property of discreteness, a number of other properties follow. For example, discrete representations should be more categorical, more easily verbalize, better for perceiving or performing a sequence of steps, and more accessible to awareness. All these properties have component of the proposed research is to explore the idea that they all follow from the more basic property of dicreteness. The initial goal of the project will be establish through experimental studies the validity of the hypothesis that working memory representations are indeed discrete. This will be done by exploring a key behavioral consequence of this hypothesis: working memory representations should be more categorical than other representations. This predicition will be investigated with a set of existing empirical paradigms that have elucidated variables that affect the working memory system, including developmental age, delay, dual-task demands, and brain damage. Working memory plays a central role in most accounts of complex human behavior, because working memory is required in any task that involves multiple stos ir a temporally extended focus of attention. This kinds of tasks are important and pervasive throughout society, including: economic, political, and military planning; air-traffic control; and scientific research, to name just a few. It is essential to understand the nature of the representations in the working memory system and to understand how people learn to use working memory in the service of complex behavior. This research will advance our knowledge in this important area, and may provide insight into techniques for rehabilitation of working memory following brain injury and techniques for assisting the development of working memory in children.

Human vision is exceptionally flexible. Consider searching for a familiar face in a crowd, finding car keys in a cluttered kitchen, chasing down an opponent in football, or threading through a crowded restaurant to find a table. How does vision accommodate such a variety of visual environments and achieve such diverse goals? One key is selective attention, which allows a person to focus on relevant aspects of the visual environment. But how does the mind determine what is relevant? What is the nature of control of visual attention?
With support from the National Science Foundation, Dr. Shaun Vecera and Dr. Michael Mozer propose research to investigate selective attention. Dr. Vecera will conduct experimental studies that explore human behavior in novel tasks and unfamiliar environments to investigate how people exploit visual cues in the environment to enhance performance. Dr. Mozer will build computer simulation models to explain the experimental results. Broader impacts of this research include building artificial systems with the flexibility of human vision and improving the design of user interfaces for computer systems and mechanical or electronic devices. Additionally the proposed research may inform the diagnosis of control deficits due to brain damage. A computer model of human control can be damaged in various ways, to simulate brain damage, leading to a better understanding of the consequences and strategies for remediation.

The proposed work will develop, implement, and evaluate new techniques that help to automate performance analysis of modern software systems. The methodology pursued breaks down the problem of automating performance analysis into three components: identifying performance anomalies, detecting covariation between performance metrics, and determining causality between covariant metrics. The proposed approach uses statistical data mining and machine learning techniques to automate the three components. The proposed system works on a collection of traces, with each trace containing one or more streams of measurements from a performance metric.

The project will bring techniques from the statistical data mining and machine learning techniques to bear on the problem of automating performance analysis. The fundamental insight of the proposed approach is that there is significant information in the time-varying contours of a stream. Previous statistical approaches to performance analysis have ignored this information in favor of examining the covariation across metrics at a particular snapshot of time.

This research project will develop an exploratory experimentation methodology for human behavioral research that will allow cognitive scientists to efficiently identify optimal conditions -- those leading to the most robust learning, the fastest performance, the fewest errors, the best decisions and choices. The tools to be developed will allow scientists to answer questions they cannot currently address due to the massive data collection effort required. To understand and predict human behavior, scientists typically perform controlled experiments that compare a small, carefully chosen set of experimental conditions. For example, in designing instructional software, a comparison might be made between two techniques for teaching students. The finding that one technique obtains reliably better outcomes has both practical and theoretical implications. However, this result does not answer the question one often wishes to ask: what is the very best possible technique? The methodology to be developed will allow scientists to evaluate many experimental conditions with only a few participants, in contrast to the traditional controlled experiment which evaluates only a few conditions each with many participants. A key product of the project will be black-box software that researchers in various disciplines of the cognitive sciences can use to apply exploratory experimentation to problems in their own field. Experimental studies also will be conducted to demonstrate the breadth of the approach in domains including: concept acquisition, color aesthetics, formal instruction, and the design of usable and engaging software.

The project will extend Bayesian optimization methods to human experimental research. Bayesian optimization has long been used in the geostatistics community for inferring unobserved properties (e.g., oil reserves below the earth's surface) from costly measurements (e.g., drilling tests). In the current project, the "landscapes" being explored are defined over possible conditions (e.g., training strategies), the unobserved properties are internal cognitive states of the human observer, and the measurements are obtained via behavioral evaluations (e.g., assessments of learning). To apply Bayesian optimization methods to a range of human experimental research, mathematical models will be developed for multiple behavioral response measures, including choice, ranking, rating, latency, and free recall. The exploratory nature of the approach requires heuristics for sequentially selecting experimental conditions to obtain maximally informative data given prior observations. Various heuristics will be evaluated in the context of behavioral research.

While traditional textbooks are designed to transmit information from the printed page to the learner, contemporary digital textbooks offer the opportunity to study learners as they interpret and process information being read. With a better understanding of a learner's state of mind, textbooks can make personalized recommendations for further study and review. How can the learner's state of mind be determined? Open a used printed textbook and the answer is clear: students feel compelled to engage with their texts by annotating key passages with highlights, tags, questions, and notes. Despite students' spontaneous desire to annotate as they read, this form of interaction has reaped few educational benefits in the past. At best, highlighted passages are re-read to study for exams, a strategy not nearly as effective as other strategies such as self-quizzing. This project will develop a new methodology that: assesses student knowledge level automatically based on annotations, transforms highlighted passages into appropriate study questions, and provides each student with well-timed, personalized review. Because the project is based on free, peer-reviewed, openly licensed materials from OpenStax that have been widely adopted at a range of institutions, particularly community colleges, the technology will reach beyond elite institutions to provide a broad spectrum of underserved students with access to a potentially powerful learning tool.

This project adopts a big-data approach that involves collecting annotations from a population of learners to draw inferences about individual learners. The project will determine how to exploit these data to model cognitive state, enabling the team to infer students' depth of understanding of facts and concepts, predict subsequent test performance, and perform interventions that improve learning outcomes. A tool will be developed that administers appropriately timed quizzes on material related to a student's highlights. A collaborative-filtering methodology will be employed that leverages population data to suggest specific passages for an individual to review. The proposed tool will reformulate selected passages into review questions that encourage the active reconstruction and elaboration of knowledge. The design and implementation of the tool will be informed by both randomized controlled studies within the innovative OpenStax textbook platform and coordinated laboratory studies. These studies will address basic scientific questions pertaining to why students annotate, how to improve their annotation skills, and techniques to optimize the use of annotations for guiding active review.