The A. Richard Newton Breakthrough Research RFP AwardsMicrosoft Research is pleased to announce the 10 recipients of the A. Richard Newton Breakthrough Research RFP Awards, totaling $1,000,000 (USD) in funding. This focus of this RFP is the identification of promising new avenues of research for solving today’s most important problems. The intent is to fund projects which clearly leverage innovative computational techniques and advantages at the heart of the research, and which may integrate approaches from a number of disciplines. Breakthrough projects are those which demonstrate potentially high impact by solving problems of great importance to science or society, or by inventing promising new approaches for enabling their solution. The A. Richard Newton Breakthrough Research RFP 2007 Award Recipients
Visualizing Voice
Karrie Karahalios
University of Illinois Audio, speech, and voice have not been given the same emphasis in the areas of human computer interaction and social visualization as have graphics and text. Speech research has been conducted by a minority and has not reached its full potential. Today, the prevalence of Voice over IP with systems such as Skype provides the next wave of speech interfaces that are being adopted by the critical masses. Using voice to communicate is easier than typing because we have thousands of years of experience talking. It is less invasive than vision (although not completely private), and computationally visualizing voice can provide many social cues and feedback that are not easily perceived in traditional face-to-face interaction. This proposal outlines a challenging goal of combining the ease of voice with the visual feedback of graphics to create a new communication medium. In a sense we are creating a graphical language for visualizing communication. We aim to use different versions of this graphical language for mediating remote and co-located conversation, for creating learning tools for acquiring new language skills and conducting speech therapy, for creating new visualization techniques combining time and phase analysis, and for novel methods of archiving audio, speech, and voice. 
Non-inductive Methodologies for Learning with Sparse Heterogeneous Data
Vladimir Cherkassky
University of Minnesota Many challenging data mining applications involve learning with sparse and heterogeneous data. In biomedical applications, there is a need to combine/utilize patients’ clinical, genomic, and demographic data for medical diagnosis. Most existing learning methods developed in statistics, machine learning, and pattern recognition are based on standard inductive learning formulations, where the goal is to estimate a predictive model from finite training data. In spite of their success, these inductive methods (such as neural networks and support vector machine methods) require significant modifications and/or clever pre-processing in dealing with sparse heterogeneous data. The proposed research will investigate several emerging non-inductive learning settings for learning with sparse heterogeneous data, including the formal (mathematical) problem statement, development of appropriate learning algorithms (including model selection) and several real-life applications. Alternative learning formulations include: Learning with Structured Data, Multi-Task Learning (MTL), and Learning through Contradictions. The conceptual aspects of these new approaches will be explored by contrasting non-inductive learning formulations with the standard inductive approach. The methodological goal of this study is the improved understanding of non-inductive learning approaches using data sets for AD diagnosis collected at the Mayo Clinic Alzheimer’s Disease Research Center. Exploring the Uncanny Valley
Jessica Hodgins, et al.
Carnegie Mellon University We propose to experimentally explore the Uncanny Valley hypothesis using eye tracking and fMRI. The stimuli will be animated sequences, modified video footage, and sequences from commercially released movies. The results of these experiments should provide a greater scientific understanding of the perception of human motion as well as providing guidelines for the production of animated human characters for use in video games, movies, and as avatars. 
The Stochastic Model Builder Applied to Single Cell Kinetics
Eric Klavins
University of Washington Stochastic processes are at the heart of many phenomena, especially those at nanoscale. Experimentalists are increasingly producing large amounts of time-series data from, for example, experiments on single genetically engineered cells. In these systems, a stochastic model is desired that predicts the time evolution of, for example, a synthetic genetic regulatory network. We propose to implement a new algorithm for building stochastic models (the Stochastic Model Builder) to solve the problem of characterizing genetically engineered systems in bacteria. The proposal involves (a) the development of software for capturing single cell behaviors from a flow chamber viewed under a microscope and (b) the development of software that finds optimal stochastic models that predict the dynamic behavior of cells in these conditions. 
Multi-scale Simulations of the Soft Elasticity of Stem Cells and Cytoskeleton/Focal-adhesion Contact
Shaofan Li
University of California, Berkeley A major advance in cellular and molecular biology is the discovery that the behaviors of stem cells depend sensitively on both the rigidity as well as the surface micro-structures of the extracellular environment. The ability of the cell to sense the environmental mechanical stimulus and subsequently to mediate its own coordinated responses is called mechanotransduction. As a process of cellular signal transduction in response to mechanical stimuli, mechanotransduction plays an important role in normal physiological processes such as cell motility, angiogenesis, embryonic development, tissue regeneration, and wound healing. The exact molecular mechanism for the mechanotransduction of focal adhesion is still unknown, and it is under active investigation. The objectives of this project are (1) to establish a predictive modeling paradigm that can help us to understand the biomechanics and biophysics underlying focal adhesion based mechanotransduction, and (2) to explain and to elucidate protein conformational changes and binding affinity changes in response to external forces, external ligand perturbations, and properties of the external environment. We propose a soft elasticity coarse-grained model that combines the strength of different modeling methods including molecular dynamics and the finite element method of contact mechanics to implement this project. The results from this study will offer insights about the role of soft elasticity on the mechanotransduction of focal adhesion, helping to unravel its molecular mechanisms on diseases, and to assist the design of targeting therapies. 
A Computational Model to Characterize the Effects of Brain Damage and to Plan Rehabilitation
Stephen Kosslyn
Harvard University Bad diet, bad genes, and bad luck can all contribute to brain damage. Stroke and closed-head injury (such as occurs in many auto accidents) are common causes of brain damage, but so are degenerative diseases (such as Alzheimer’s). Researchers and physicians are very good at assessing which part or parts of the brain have been physically damaged, but not very good at quickly and precisely diagnosing what the consequences of the damage will be on a person's thinking, emotions, and behavior. This research will have a profound effect on neurology, and possibly have implications for artificial intelligence. Our efforts to understand the effects of brain damage cannot help but illuminate facts about the normal brain. Our proposed project will be unique. We will build a computer simulation model to: (i) emulate distinct functions within the context of a single integrated dynamic network; (ii) exploit Bayesian statistics to make use of stored information to regulate network activity; and (iii) mimic the dynamic effects of damaging either a function or connections among functions. In addition, we will (iv) develop tasks that can be used clinically with humans, allowing a direct correspondence between deficits exhibited by the simulation model and those that can be documented in humans, and (v) begin to use the model to predict the best regimen of rehabilitation for a specific patient. 
Faults, Bugs, IP Protection, and Secure Hardware
Igor Markov
University of Michigan Ann Arbor This proposal addresses several rising challenges that are starting to affect the electronics and software industries. One is the fact that sophisticated chips are almost impossible to debug by existing techniques before they are manufactured, at least not within time-to-market constraints. Therefore, computer-aided debugging at different stages is becoming increasingly necessary. The second challenge – intellectual property protection – is magnified by the outsourcing of manufacturing into countries with loose enforcement of IP rights. We propose new techniques for IP protection based on deliberately planting bugs or design faults into integrated circuits (ICs) in such a way that these bugs can be removed by entering an activation key that is different for each chip. These techniques will also improve software IP protection by providing a trusted hardware platform. The third challenge, recently articulated by the Defense Science Board, is the development of trusted hardware. Here our main contributions are to articulate the difficulty of the problem by developing simple techniques for subverting cryptography hardware. These techniques are based on doctored logic synthesis tools and appear to be very difficult to counter-act. We discuss several partial solutions related to tamper-proof chips, which help with only some aspects of the problem. A more profound way of addressing this problem is by leveraging signal observability during verification and chip testing. 
Parallel Large-scale Semi-definite Programming for Molecular Electronic Structure
David Mazziotti
The University of Chicago In 1959, Charles Coulson proposed banishing the many-electron wave function from molecular computations in favor of a two-electron density. Our research group recently made a critical breakthrough for realizing Coulson's dream of computing quantum properties of many-electron atoms and molecules from only two electrons. We plan to develop a large-scale parallel two-electron algorithm to study an array of many-electron phenomena, difficult or impossible to treat with approximate wave function methods, from chemical reactivity to superconductivity. Research will impact the frontiers of chemistry and condensed-matter physics by complementing experimental advances. Within the two-electron approach a special type of constrained optimization known as semi-definite programming minimizes the ground-state quantum energy. The development of an efficient, parallel algorithm for semi-definite programming will have broad applications in diverse scientific areas from combinatorial optimization, control theory, and economics to number theory, quantum information, and quantum computing. 
How to Build a Habitable Planet: Estimating the Physics of Plate-tectonic Convection on Earth
Jun Korenaga
Yale University Understanding the physics of plate-tectonic convection in Earth's mantle is one of the outstanding and most puzzling challenges in the geosciences and planetary sciences, with profound implications for the habitability of a terrestrial planet and the evolution of life. The strength of mantle materials is strongly temperature-dependent, which should prevent the operation of plate tectonics, so there must be some weakening mechanism to compensate, but what this mechanism could be is currently unresolved. This project will build a comprehensive statistical framework to constrain macroscopic mantle properties using geophysical observations and computational fluid dynamics. This project represents a major step toward a self-consistent theory of plate tectonic convection, and exemplifies how we can approach this long-standing mystery by addressing a fundamental physics question and formulating it as a quantitative mathematical problem, the solution of which has finally become tractable owning to the rapid development of computational technology. 
Integrated Probabilistic Models of Regulatory and Metabolic Networks for Bio-fuels Research
Yuan Qi
Purdue University Developing alternative renewable energy sources such as bio-fuels, which can be derived from the cellulose in plant cell walls, is crucial for the sustainable growth of human societies. To genetically improve energy crops and efficiently convert biomass to cellulosic bio-fuels, we need to answer basic scientific questions regarding gene regulation and metabolic pathways in plant cells, so that we can breed plants that have optimal energy conversion characteristics.
There is a critical need for new computational strategies that allow us to extract knowledge from heterogeneous data sets, such as genomic sequences, expression arrays, and metabolite flux analyses, and to integrate different biological network models to answer the crucial scientific questions that will make bio-fuel production a reality.
Our long term goal is to develop principled, efficient computational and statistical tools to construct joint network models for transcriptional regulation and metabolic pathways from heterogeneous data sources, and utilize the integrated network models to predict target phenotypes. We will develop a unified probabilistic network framework that not only models the combinatory regulation of transcription factors, but also integrates gene regulation with physico-chemical constraints in metabolic pathways.
Our specific aims include 1) identifying metabolic genes, especially, transcription factors regulating lignin biosynthesis, 2) building regulatory networks regulating lignin biosynthesis, and 3) integrating the regulatory and metabolic networks to predict gene expression and metabolite concentrations, e.g., lignin yields. The reconstructed, integrated network model and its computational predictions will be evaluated by both genetic and biochemical assays. The completion of this project will provide the bio-fuels research community with important knowledge on the gene regulation of lignin biosynthesis, paving the way for engineering energy crops with low amounts of, or easily removable lignin. The A. Richard Newton Breakthrough Research Award Request for Proposals 2007 RFP
|
