Yoav Freund - US grants

Large-Scale numerical simulations demand scalable solvers that strike a delicate balance between robustness and cost. The proposed project aims to solve the complex problems of algorithm selection and tuning by implementing a heuristic decision-making agent. Through techniques from data mining and machine learning, applied to production runs, the decision-making component will increase its repertoire of heuristics and tune these heuristics. This will enable many areas of computational science and engineering to use sophisticated numerical software with high efficiency, profoundly reducing the execution time of production runs. We will work on solver software in wide use, over which we have source code control to permit necessary monitoring and manipulation.


The innovative aspects of this project are in the integration of numerical analysis and techniques from mainstream computer science such as machine learning, data mining, and componentization, creating a software architecture that makes it easy to employ numerical techniques in an application context with expert knowledge, and to codify their benefits for wider community.

This project studies machine learning from data that appears high dimensional but, in fact, has low intrinsic dimension (e.g., the data lies on a low-dimensional manifold). Physical constraints in many applications produce exactly such a situation. The project is developing machine learning systems that use resources (e.g., compuational time and space) that scale with the intrinsic rather than the extrinsic dimension. The idea of data lying on a manifold is appealing and suggestive, and has been the inspiration of a lot of recent, exciting work in machine learning. Often the aim is to embed such data into a lower dimensional space, after which the application of standard methods consume less resources. The PIs have developed a precise notion of intrinsic dimension that captures the manifold intuition while being broad enough to both be statistically sensible and empirically verifiable. This quantity is then treated as a fundamental parameter in terms of which a variety of new nonparametric methods can be assessed. The first of these is a simple variant of the k-d tree that is provably adaptive to intrinsic dimension. The PIs also consider schemes for nonparametric classification and regression, for manifold learning, and for embedding. These new algorithms and ideas will be applied to fundamental challenges in a variety of domains, including sensor networks, computer vision, protein structure prediction, and robotic control.

The ease of use of a human computer interface depends critically on the latency of the system and its ability to adapt to the environment. Differences in lighting, visual appearance and user behavior can significantly alter the input data. Furthermore, the reaction time must be less than 100 milliseconds to appear instantaneous to the user. Achieving high accuracy demands a system that adapts to these changing characteristics while processing a significant amount of data in a short amount of time.

We propose a computer architecture for adaptive real-time signal processing systems that combines a general purpose processor with custom hardware. The custom hardware performs the low-level, high-throughput signal processing on the raw signals and feeds them to the processor which performs the high level signal processing and decision making. The processor also executes machine learning algorithms that change the parameters of the low-level processing to adapt them to the current statistical properties of the data.

This project will develop a human-computer interface based on audio and video sensors that allows a user to interact with the computer through gestures and voice alone. This requires research advances in computer architecture, embedded systems, signal processing, machine learning and human-computer interaction. The major research challenge is in the integration of knowledge from the different areas to create a functional system. This system will serve as a prototype for novel human computer interactions and will be a foundation for future collaboration between the different fields.

This project defines meaningful notions of confidence in prediction, designs procedures for computing such notions, and applies these procedures to core machine learning tasks such as active learning, crowd-sourced learning, and tracking. In many applications it is helpful to have classifiers that output, together with each prediction, a rating of the confidence that the prediction is in fact correct. Existing literature either provides various ad-hoc ways for computing such ratings which typically lack a rigorous mathematical footing, or provides mathematically consistent methods (in the Bayesian framework) for computing confidence ratings under very strong assumptions that are unlikely to hold in practice. The research team investigates methods of computing measures of confidence that are mathematically rigorous while making minimal assumptions on the way data is generated, and use these measures to further develop solutions to core machine learning tasks.

Defining and computing mathematically sound measures of confidence lies at the heart of machine learning, pattern recognition and uncertainty in AI. Confidence-rated prediction, active learning, and tracking are fundamental tasks of machine learning and statistics that arise repeatedly in large-scale problems; this project will develop rigorous solutions to these problems. The algorithms developed in this work are tested and used in the Automatic Cameraman project, an interactive, audio-visual installation in the UCSD Computer Science department. The interactive Automatic Cameraman system are used an educational tool to be extended in many different directions, by teams of students at a variety of skill levels.

Sensor data of diverse types and large volumes need to be combined with the current standard SQL databases, which provide context and metadata for the sensor data. The combination will lead to a new generation of analytics in a number of areas, such as smart buildings that are based on building and environmental data collected by sensors. The project argues that this new generation of analytics must be based on the same healthy database technology cornerstones that the prior (non-sensor) business intelligence platforms were based on: Declarative queries, automatic optimization, efficient storage representations and multiple layers of abstraction lead to high productivity for the developer and the analyst. Such productivity is currently absent from sensor data analytics because database technology and sensor data processing currently do not mix well. Productivity is especially low in cases involving (a) many types of sensor data, (b) combinations of sensor data with conventional database data that provide context and (c) many types of analyses. Besides low productivity, the current (limited) state of the art poses very high expertise requirements on the analysts: They must be simultaneously experts in signal processing, statistics and big data management. The project will deliver a database system for sensor data, where the analyst can rapidly develop declarative queries that are automatically optimized. By doing so, the project will deliver the envisioned productivity gains and will lower the technical sophistication bar needed for acting in the space, therefore enabling many scientists and domain specialists to engage in analytics.

This project argues that at the core of the failure of SQL databases in the management and analytics of sensor spatiotemporal data is the lack of a critical abstraction, which is the real world models, which capture the stochastic processes that generate the measurements. The proposed Plato database system will bring the real world model concept into SQL databases by using models (spatiotemporal continuous functions) as first class citizens. The delivery of Plato requires innovative solutions to multiple problems: The project will design and
implement (a) a model-aware data model and respective query language features that allow seamless combination of conventional SQL querying with statistical signal processing, (b) learning algorithms that learn the model components of reduced-noise, additive model representations, which are naturally compressions of the original, (c) query processing algorithms that operate directly on the compressed representations and utilize the the relatively few bits necessary for the required confidence of the analytics, and (d) semiautomated algorithms that further compress the model representations by considering the dependencies (mutual entropy) between the models. Finally, the project will exercise the resulting system on large scale statistical sensor data processing cases, such as the ones presented by the UCSD Energy Dashboard. The exercise will measure the lines-of-code as well as the runtime efficiency of the analyses.

For further information see the project web site at http://www.db.ucsd.edu/NSF14Plato

? DESCRIPTION (provided by applicant): Neuronal circuits in the brainstem control life-sustaining functions, in addition to driving and gating active sensation through taste, smell, and touch. We propose to exploit the advent of molecular and genetic tools to undertake cell lineage marking, cell phenotyping, molecular connectomics, and methods from machine learning and image processing to construct an integrated anatomical and functional atlas of the brainstem. This will enable us to generate anatomical wiring diagrams for the brainstem circuits that control or facial actions. There are three phases to this work. (1) Reveal the identity and organization of brainstem nuclei. Motivated by striking similarities between the developmental plan for the spinal cord and brainstem, we will embrace and extend these efforts to interrogate the molecular composition of neurons that define individual nuclei with sensorimotor circuits in the murine brainstem. (2) Reveal brainstem neuronal circuits and their interactions. We will utilize Tran synaptic viral labeling to delimit pathways from specific muscles that are innervated by facial, trigeminal, hypoglossal, and laryngeal motor nuclei. This will reveal hitherto unknown brainstem circuits, including sites of modulation by higher brain areas. (3) Control the behavior of identified feedback circuits. We will manipulate specific populations of brainstem neurons using a battery of genetic tools to delineate or facial motor actions and motor synergies. The results from the above efforts will be a quantitative map of the functional organization of neurons in the brainstem that enable studies on computations that underlie or facial behavior. An understanding of these fundamental behaviors bears directly on the more general issue of how nervous systems deal with computations that can be performed autonomously, yet must interact synergistically. Thus our proposed program on brainstem circuitry and dynamics will yield general lessons about the nature of neuronal computation. The work performed under this proposal will serve as the basis for a larger national effort in brainstem neuronal computation.

? DESCRIPTION (provided by applicant): Neuronal circuits in the brainstem control life-sustaining functions, in addition to driving and gating active sensation through taste, smell, and touch. We propose to exploit the advent of molecular and genetic tools to undertake cell lineage marking, cell phenotyping, molecular connectomics, and methods from machine learning and image processing to construct an integrated anatomical and functional atlas of the brainstem. This will enable us to generate anatomical wiring diagrams for the brainstem circuits that control or facial actions. There are three phases to this work. (1) Reveal the identity and organization of brainstem nuclei. Motivated by striking similarities between the developmental plan for the spinal cord and brainstem, we will embrace and extend these efforts to interrogate the molecular composition of neurons that define individual nuclei with sensorimotor circuits in the murine brainstem. (2) Reveal brainstem neuronal circuits and their interactions. We will utilize Tran synaptic viral labeling to delimit pathways from specific muscles that are innervated by facial, trigeminal, hypoglossal, and laryngeal motor nuclei. This will reveal hitherto unknown brainstem circuits, including sites of modulation by higher brain areas. (3) Control the behavior of identified feedback circuits. We will manipulate specific populations of brainstem neurons using a battery of genetic tools to delineate or facial motor actions and motor synergies. The results from the above efforts will be a quantitative map of the functional organization of neurons in the brainstem that enable studies on computations that underlie or facial behavior. An understanding of these fundamental behaviors bears directly on the more general issue of how nervous systems deal with computations that can be performed autonomously, yet must interact synergistically. Thus our proposed program on brainstem circuitry and dynamics will yield general lessons about the nature of neuronal computation. The work performed under this proposal will serve as the basis for a larger national effort in brainstem neuronal computation.

Project 5. Abstract Computational Neuroanatomy (Yoav Freund, lead; Friedman, Karten, Kleinfeld) Anatomical atlases play an essential role for characterization of circuitry by collation of ?Components? which in turn enables reverse engineering of these circuits. Control of orofacial actions is coordinated by distinct populations of brain stem premotor neurons, which are arranged into relatively small clusters and can be limited to domains as small as 200 to 300 µm in extent. Further, for many orofacial motor actions, premotor neuronal clusters are present at multiple levels of the brainstem and do not conform to the boundaries previously defined by available atlases, including the Paxinos atlases and the Allen Brain Common Coordinate Framework atlas. We propose to construct a Trainable Texture-based Digital Atlas from digitized stacks of brain images obtained by tape-transfer of serial cryosections through the brain (Core 2 - Precision Histology) to enable mapping of the brainstem premotor interface modulation of orofacial motor actions. The atlas design allows labeled cells, projections and recording sites to be accurately and automatically aligned across different brains. Our Trainable Texture-based Digital Atlas makes use of identification of landmarks based on texture features of Nissl stained cytoarchitecture. The landmarks are identified by expert anatomists and are used to create training sets for machine learning. Machine learning is used to train texture detectors to distinguish between different cytoarchitectural textures in order to automate landmark identification that is consistent with the original manual landmark annotations by anatomists. This process and the automated alignment of new brains is performed in three dimensions The Trainable Texture-based Digital Atlas is implemented on a computer cloud server (Core 3 - Data Science). This enables us to integrate experimental results across all of the project participants and data from others outside our project. Thus the Digital Atlas is platform-independent. Our data management is designed to facilitate accessibility of the atlas, of meta data that describes experimental output, and of mappings back to all slices in each brain, which is expected to take at most a few Gbytes. All users will be able to efficiently browse the Digital Atlas and meta-data. It will also be possible retrieve subsets of images from full brain stacks for validation of raw data. Our particular focus is on the brainstem. Yet the system is general and can be expanded to the entire central nervous system; indeed, a new graduate student has begun work on a joint project with Dr. Einman Azim (Salk Institute), to extend the atlas to the spinal cord, another CNS region with challenging cytoarchitectural borders for subregion parcellation.

PIs: Martin Deschênes, Yoav Freund, David Golomb, David Kleinfeld (lead), Fan Wang ! Core 3. Abstract Data Science Core The role of the Data Science Core is to maintain the digital data and to link to it all experimental results and derived information (meta-data) from all participants. Material for the Trainable Texture-based Digital Atlas is a major part of this effort. An analysis workflow is maintained that serves to localize histology stacks and map experimental results to atlas coordinates. The Core also serves to disseminate the raw and meta data and all associated software from all studies at the time of acceptance of publications.

Overview - Abstract Brainstem function is necessary for life-sustaining functions such as breathing and for survival functions, such as foraging for food. Individual motor actions are activated by specific brainstem cranial motor nuclei. The specificity of individual motor actions reflects the participation of motor nuclei in circuits within closed loops between sensors and muscle actuators. However, these loops are also nested and connect to feedback and feedforward pathways, which underlie coordination between orofacial motor actions. A key question for this proposal is how different actions are coordinated to form a rich repertoire of behaviors, such as rhythmic motions linked to breathing, and the orchestrated displacements of the head, nose, tongue, and vibrissae during exploration. We postulate that the best candidate interface for orofacial motor coordination are premotor and pre2motor neuron populations in the brainstem reticular formation: these neurons project to cranial motor nuclei, receive descending inputs from outside of the brainstem, and interconnected to each other. Our approach exploits and expands upon a broad spectrum of innovative experimental tools. These include state-of-the-art behavioral methods to study motor actions and their coordination into behaviors. From an experimental perspective, the underlying neuronal circuitry for each orofacial motor action may be accessed via transsynaptic transport starting at the muscle activators or associated sensors in the periphery. These studies will make use of molecular, genetic, and functional labeling methods to enable cell phenotyping and circuit tracing. These data will establish the Components, i.e., brainstem nuclei connectivity for all Research Projects. These studies are complemented by in vivo electrophysiology and optogenetics in order measure and perturb the signal flow during exploration and decision-making: these studies will establish orofacial ?Wiring Diagrams?. The sum of these techniques will permit us to elucidate the functions of intrinsic brainstem circuits and their modulation by descending pathways. Our data will be integrated in two ways. First we will begin development of computational models of the dynamics of active sensing by the orofacial motor plant and brainstem circuits. These will initially focus on the vibrissa system, starting with characterizations of mechanics and mechano-neuronal transformations of vibrissa movement and extending to exploration of brainstem circuits that drive vibrissa set-point and rhythmic whisking. Finally, vibrissa feedforward pathways will be computationally modeled to explore how sensory input affects vibrissa dynamics. Second, to record connectivity data that arises from our experimental tracing studies, we will construct an Trainable Texture-based Digital Atlas that utilizes machine learning to automate anatomical annotation of brainstem nuclei. The Atlas is designed to allow accurate 3D alignment of labeled neurons, even when labeled neurons reside in small sub-regions outside of well-defined brainstem nuclei, based on triangulation to Atlas landmark structures. Further, digitization of serially sectioned brain data sets allows 3D reconstruction and alignment of small brainstem subregions as well as the collation of this data from different brains into the same Atlas. Our proposed program on brainstem circuitry and dynamics will yield general lessons about the nature of neuronal computation. The analytic and anatomical tools developed for these studies will be made available through our data science core to the larger neuroscience community.