Joshua Tzvi Vogelstein - US grants

Understanding how behavior emerges from the dynamic patterns of electrical and chemical activity of brain circuits is universally recognized as a fundamental mystery in science. Furthermore, better knowledge of healthy brain function has broad societal implications by laying the groundwork for advancing treatments for neurological disorders and for developing brain-inspired technologies. Various brain research efforts around the globe have embraced this scientific grand challenge. This meeting brings together US and international brain researchers to discuss the need for coordinating brain research efforts around the globe, the bottlenecks and challenges that stand in the way and strategies to overcome these. A particular focus of the meeting is on cyberinfrastructure and data resources needed to further enhance collaboration and analysis of disparate streams of neuroscience data.

The Organizers are making a strong effort to invite women and members of underrepresented groups as participants. Further, the meeting entails extensive cross-disciplinary interactions, which will be aided greatly by the face-to-face nature of this meeting. To maximize the workshop's impact, a white paper summarizing the discussion will be published at a high-profile venue.

The brain research community needs to increase the practice of sharing and combining data sets to increase the power of statistical analyses and to gain the most knowledge from collected data. This project aims to build a prototype system called BrainLab CI that will facilitate meaningful integration of thousands of publicly available Magnetic Resonance Imaging (MRI) and neurophysiology data sets, and allow researchers to define and conduct new large-scale community-level experiments on these data. BrainLab CI has the potential to transform research practice in neuroscience by overcoming major obstacles to data sharing: Scientists will be able to share data without losing control over data quality, and will maintain full visibility into how all subsequent experiments use their data and algorithms. This project may consequently drive a change in scientific culture by encouraging data sharing and the development of common analysis tools, and resulting accelerated discovery from connecting ideas, tools, data, and people. This project therefore aligns with the NSF mission to promote the progress of science and to advance the national health, prosperity and welfare.
The BrainLab CI prototype system will provide new paradigms for combining different analytic methods, meta-analysis with raw data, comparing the results of different laboratories and even synthesizing new experiments by combining different studies. An experimental-management software system will be deployed that allows users to construct community-wide experiments that implement data and metadata controls on the inclusion and exclusion of data. Example of controls include: requiring specific metadata, that data are registered to a given atlas, or that data are collected using specific experimentation protocols. BrainLab CI will initially focus on two different experimental patterns: (1) An incremental experiment defines an experiment against an existing data set which then opens to additional community contributions of data; and (2) a derived experiment forks/branches an existing experiment, allowing a researcher to change properties, such as an acceptance criteria or analysis algorithm, but otherwise run the same pipeline against the same inputs. The system will allow each experiment to maintain online dashboards showing how additional data changes results with complete provenance. To develop and validate the BrainLab CI prototype, several community experiments will be developed for MRI and for neurophysiology (including both optical and electrical physiology) data. These research domains were chosen because of the great potential gains for increased sharing of laboratory data in these domains.
This Early-concept Grants for Exploratory Research (EAGER) award by the CISE Division of Advanced Cyberinfrastructure is jointly supported by the SBE Division of Behavioral and Cognitive Sciences, with funds associated with the NSF Understanding the Brain activity including for developing national research infrastructure for neuroscience, and alignment with NSF objectives under the National Strategic Computing Initiative.

Detecting relationships between two data sets has long been one of the most important questions in statistics and is fundamental to scientific discovery in the big-data era. By developing an open-source, robust, efficient, and scalable statistical methodology for testing dependence on modern data, this project aims to advance the understanding and utility of testing dependence, tackle a number of related statistical inference questions, and accelerate a broad range of data-intensive research. The project incorporates fundamental research in mathematics, statistics, and computer science to further develop a multiscale generalized correlation framework to enable discovery and decision-making via analysis of large and complex data. The tools under development will allow scientists to better explore and understand high-dimensional, nonlinear, and multi-modal data in a myriad of applications. The project aims to provide a unified framework for discovery of relationships between observations in an efficient and theoretically-sound manner.

Combining the notion of generalized correlation with the locality principle, multiscale generalized correlation (MGC) is a superior correlation measure that equals the optimal local correlation among all possible local scales. By building upon distance correlation and making use of nearest neighbors, the resulting MGC test statistic is a unique dependence measure that is consistent for testing against all dependencies with finite second moment, and it exhibits better performance than existing state-of-art methods under a wide variety of nonlinear and high-dimensional dependencies. By investigating the theoretical aspects of distance-based correlations, this project aims to further improve the finite-sample performance of MGC-style tests, extend its capability to testing dependence on network and kernel data, and broaden its utility to general inferential questions beyond dependence testing such as two-sample testing, outlier detection, and feature screening, as well as applications to brain activity, networks, and text analysis. Overall, this project intends to establish a unified methodology framework for statistical testing in high-dimensional, noisy, big data, through theoretical advancements, comprehensive simulations, and real data experiments.

Three complimentary changes are revolutionizing the way neuroscientists study the brain. First, experimental advances allow neurobiologists to "clear" brains so that they become transparent, with the exception of a set of neurons that can be selected on the basis of their location, response properties, and genetic make-up. Second, technological advances have resulted in microscopes that can simultaneously image an entire "sheet" of this brain, thereby enabling rapid acquisition of whole brain volumes. Third, researchers are taking steps to educate neuroscientists to acquire these data. Together, this will result in a massive upswing in adoption of this experimental modality. However, acquiring the data is one step in the upward spiral of science that will yield transformative scientific results. The subsequent steps are computational. This project will develop cyberinfrastructure resources and software that enable storage and access of large CLARITY brain imaging datasets, alignment and registration to reference anatomical atlas and visualization of the datasets. Additional capabilities for automatic identification and localization of cell bodies and statistical analysis will be provided. The PIs will run annual hackathons for college students and sponsor a summer internship program for undergraduates to broaden the educational efforts in software development for neuroscience. Finally, mobile compliant digital education content will be created to complement existing online courses to target STEM students, and educate global citizens.

This project will build a prototype pipeline that operates on raw CLARITY brains and outputs the statistics of locations of cells in each region in the Allen Reference Atlas, as well as estimates of connectivity and similarity across regions and conditioned on different contexts. To do so, the PIs will leverage modern mathematical statistics (such as Large Deformation Diffeomorphic Metric Mapping for registration, Deep Learning and Random Forests for segmentation, and Statistical Graph Theory for analysis of the resulting conenctomes), as well as modern computational tools, including Docker containers to facilitate full reproducability, and semi-external memory algorithms and cloud computing to enable scalable analytics. To reach out to the broader community and educate them in the use of these tools, this project will provide tutorials deployed in the cloud. Together, this will facilitate the large community of users to both collect and analyze their data with ease. Many of the tools developed as part of this project will be easily extensible to other experimental modalities and neuroscience communities.

Large, three-dimensional cell colonies grown inexpensively using simple raw materials could be made into cheap, energy-efficient computers. A fundamental challenge in using living cells for computing is that computation by cells is error prone, and cells divide, die and reorganize inside a cell culture, making it difficult to maintain a defined architecture. This research will explore the design of yeast cell-based computing systems inspired by how computing is performed by the animal brain cells. To develop new knowledge at the intersection of electronics, computing and biology will require a new generation of students familiar with each of these areas who can work in collaborative teams. Building on work with organizations including the Freshman Research Initiative at UT Austin and Women in Science and Engineering at JHU, the PIs will develop programs to allow groups of undergraduate researchers to engage in long term research programs in which students have the opportunity to perform independent investigations as part of collaborative, inter-university teams.

This project will combine ideas from computer architecture and systems neuroscience with new tools from synthetic biology to develop yeastons - Saccharomyces cerevisiae cells that can collectively emulate a feedforward neural network through engineered cell-cell communication processes and programmable transcriptional logic. Crucially, yeaston networks will be designed to tolerate the inherent noisiness of single-cell biomolecular information processing and require no specific higher order spatial organization or patterning. The project members will build new protein receptors for small molecule signals and genetic logic systems that will enable single yeastons to emulate nodes in a feedforward neural network. The input-output behavior of single yeastons and yeaston networks will be quantitatively characterized, making it possible to evaluate the potential for scalable computation in yeaston systems. High-level models from neuroscience will be used to develop design principles for assembling robust yeaston networks and to derive scaling laws for yeaston computing.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.

Recent advances in deep learning have led to many disruptive technologies: from automatic speech recognition systems, to automated supermarkets, to self-driving cars. However, the complex and large-scale nature of deep networks makes them hard to analyze and, therefore, they are mostly used as black-boxes without formal guarantees on their performance. For example, deep networks provide a self-reported confidence score, but they are frequently inaccurate and uncalibrated, or likely to make large mistakes on rare cases. Moreover, the design of deep networks remains an art and is largely driven by empirical performance on a dataset. As deep learning systems are increasingly employed in our daily lives, it becomes critical to understand if their predictions satisfy certain desired properties. The goal of this NSF-Simons Research Collaboration on the Mathematical and Scientific Foundations of Deep Learning is to develop a mathematical, statistical and computational framework that helps explain the success of current network architectures, understand its pitfalls, and guide the design of novel architectures with guaranteed confidence, robustness, interpretability, optimality, and transferability. This project will train a diverse STEM workforce with data science skills that are essential for the global competitiveness of the US economy by creating new undergraduate and graduate programs in the foundations of data science and organizing a series of collaborative research events, including semester research programs and summer schools on the foundations of deep learning. This project will also impact women and underrepresented minorities by involving undergraduates in the foundations of data science.

Deep networks have led to dramatic improvements in the performance of pattern recognition systems. However, the mathematical reasons for this success remain elusive. For instance, it is not clear why deep networks generalize or transfer to new tasks, or why simple optimization strategies can reach a local or global minimum of the associated non-convex optimization problem. Moreover, there is no principled way of designing the architecture of the network so that it satisfies certain desired properties, such as expressivity, transferability, optimality and robustness. This project brings together a multidisciplinary team of mathematicians, statisticians, theoretical computer scientists, and electrical engineers to develop the mathematical and scientific foundations of deep learning. The project is divided in four main thrusts. The analysis thrust will use principles from approximation theory, information theory, statistical inference, and robust control to analyze properties of deep networks such as expressivity, interpretability, confidence, fairness and robustness. The learning thrust will use principles from dynamical systems, non-convex and stochastic optimization, statistical learning theory, adaptive control, and high-dimensional statistics to design and analyze learning algorithms with guaranteed convergence, optimality and generalization properties. The design thrust will use principles from algebra, geometry, topology, graph theory and optimization to design and learn network architectures that capture algebraic, geometric and graph structures in both the data and the task. The transferability thrust will use principles from multiscale analysis and modeling, reinforcement learning, and Markov decision processes to design and study data representations that are suitable for learning from and transferring to multiple tasks.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.

PROJECT SUMMARY Overview: We will extend and develop implementations of foundational methods for analyzing populations of attributed connectomes. Our toolbox will enable brain scientists to (1) infer latent structure from individual connectomes, (2) identify meaningful clusters among populations of connectomes, and (3) detect relationships between connectomes and multivariate phenotypes. The methods we develop and extend will naturally overcome the challenges inherent in connectomics: high-dimensional non-Euclidean data with multi-level nonlinear interactions. Our implementations will comply with the highest open-source standards by: providing extensive online documentation and extended tutorials, hosting workshops to demonstrate our tools on an annual basis, and merging our implementations into commonly used packages such as scikit-learn [1], scipy [2], and networkx [3]. All of the code we develop is open source. We strive to ensure that our code is shared in accordance with the strictest guiding principles. We chose to implement these algorithms in Python due to its wide adoption in the neuroscience and data science fields. In particular, many other neuroscience tools applicable to connectomics, including NetworkX DiPy, mindboggle, nilearn, and nipy, are also implemented in Python. This will enable researchers to chain our analysis tools onto pre-existing pipelines for data preprocessing and visualization. Nonetheless, we feel that sharing our code in our own public repositories is insufficient for global reach. We have also begun reaching out to developers of the leading data science packages in python, including scipy, sklearn, networkx, scikit-image, and DiPy. For each of those packages, we have informal approval to begin integrating algorithms that we have developed. Those packages are collectively used by >220,000 other packages, so merging our algorithms into those packages will significantly extend our global reach. All researchers investigating connectomics, including all the authors of the 24,000 papers that mention the word ?connectome?, will be able to apply state-of-the-art statistical theory and methods to their data. Currently, we have about 150 open source software projects on our NeuroData GitHub organization. Collectively, these projects get about 2,000 downloads and >11,000 views per month. As we incorporate additional functionality as described in this proposal, we expect far more researchers across disciplines and sectors will utilize our software. 20 ? ?? ? ??