| Date: | September 10, 2019 |
| Time: | 5:20pm |
| Location: | BLOC 220 |
| Speaker: | Boris Hanin, Texas A&M |
| Title: | Hyperplane Arrangements in Deep Learning |
| Abstract: | Deep learning is the study and application of neural networks. Each network is a non-linear family of functions, and such families form the backbone for many state of the art machine learning tasks ranging from computer vision to natural language processing. I will explain what a neural network is and will focus on a simple but important example called ReLU networks. The complexity of the function computed by a ReLU network is naturally captured by a generalization of a hyperplane arrangement. I will discuss this connection and will explain why the complexity of such arrangements is important for understanding how and why neural networks work in practice. |
|
| Date: | September 26, 2019 |
| Time: | 5:15pm |
| Location: | Bloc 220 |
| Speaker: | Weston Baines, Texas A&M |
| Title: | Filtering Shot Noise in Lidar Data |
| Abstract: | Light Detection and Ranging (Lidar) is a remote sensing method used to
produce 3D terrain maps.
Lidar scans are expensive and time consuming, sometimes requiring
several passes of a scene to obtain acceptable data fidelity.
For a low number of scene passes lidar data is primarily corrupted by
shot noise. It is desirable to filter the shot noise to increase data
efficiency.
In this talk I will discuss the basic operation of lidar systems and the
work I did at the Geospatial Research Laboratory this past summer in
removing shot noise from Lidar data. |
|
| Date: | October 14, 2019 |
| Time: | 11:00am |
| Location: | BLOC 220 |
| Speaker: | TBA, Enthought |
| Title: | Internship at Enthought |
| Abstract: | A presentation to learn more about Enthought, and it’s summer internship program.
Founded in 2001, Enthought helped establish Python's scientific community and became an
early leader in scientific digital transformation. Our team of “scientists who code” pairs in-depth
science expertise with data strategy, modeling, simulation, AI and more to transform businesses
that depend on science. We solve complex problems for some of the most innovative and
respected organizations across the oil and gas, life sciences, chemical, and semiconductor
industries. |
|
| Date: | November 7, 2019 |
| Time: | 5:00pm |
| Location: | BLOCK 220 |
| Speaker: | Hittinger, Jeffrey A. F., Lawrence Livermore National Laboratory |
| Title: | Variable Precision Computing Research in the Center for Applied Scientific Computing |
| Abstract: | In collaboration with academic, industrial, and other government laboratory partners, Lawrence Livermore National Laboratory’s Center for Applied Scientific Computing (CASC) conducts world-class scientific research and development on problems in computer science, computational physics, applied mathematics, and data science. CASC applies the power of high-performance computing and the efficiency of modern computational methods to the realms of stockpile stewardship, cyber and energy security, knowledge discovery for intelligence applications, and basic scientific discovery. As the focus for research efforts in the Computation Directorate, CASC also leads the development of methods and techniques that advance the discipline of scientific computing.
In CASC, we are developing the methods and tools that will enable the routine use of dynamically adjustable precision at a per-bit level depending on the needs of the task at hand. We typically compute and store simulation results in 64-bit double precision by default, even when very few significant digits are meaningful. Many of these bits represent errors – truncation, iteration, roundoff – instead of useful information about the computed solution. This over-allocation of resources is wasteful of power, bandwidth, storage, and operations. Just as adaptive mesh refinement frameworks adapt spatial grid resolution to the needs of the underlying solution, our goal is to provide more or less precision as needed locally. Acceptance from the community will require that we address three concerns: that we can ensure accuracy, ensure efficiency, and ensure ease of use in development, debugging, and application.
In this talk, I will present an overview of CASC and then take a deeper dive into the benefits and the challenges of variable precision computing, highlighting aspects of our ongoing research in data representations, numerical algorithms, and testing and development tools. |
|
| Date: | November 18, 2019 |
| Time: | 4:00pm |
| Location: | Bloc220 |
| Speaker: | Dr. Nikki Meshkat, Santa Clara University |
| Title: | Structural Identifiability of Biological Models |
| Abstract: | Parameter identifiability analysis addresses the problem of which unknown parameters of a model can be determined from given input/output data. If all of the parameters of a model can be determined from data, the parameters and the model are called identifiable. However, if some subset of the parameters can not be determined from data, the model is called unidentifiable. We examine this problem for the case of perfect input/output data, i.e. absent of any experimental noise. This is called the structural identifiability problem. We show that, even in the ideal case of perfect input/output data, many biological models are structurally unidentifiable, meaning some subset of the parameters can take on an infinite number of values, yet yield the same input/output data. In this case, one attempts to reparametrize the model in terms of new parameters that can be determined from data. In this talk, we discuss the problem of finding an identifiable reparametrization and give necessary and sufficient conditions for a certain class of linear compartmental models to have an identifiable reparametrization. We also discuss finding classes of identifiable models and finding identifiable submodels of identifiable models. Our work uses graph theory and tools from computational algebra. This is joint work with Elizabeth Gross and Anne Shiu.
|
|
Menu