Session L61: Data Science Platforms: Algorithms and Visualization

Live

We have witnessed tremendous growth in Artificial intelligence (AI) and machine learning (ML) recently. However, research shows that AI and ML models are often vulnerable to adversarial attacks, and their predictions can be difficult to understand, evaluate and ultimately act upon.

Discovering real-world vulnerabilities of deep neural networks and countermeasures to mitigate such threats has become essential to successful deployment of AI in security settings. We present our joint works with Intel which include the first targeted physical adversarial attack (ShapeShifter) that fools state-of-the-art object detectors; a fast defense (SHIELD) that removes digital adversarial noise by stochastic data compression; and interactive systems (ADAGIO and MLsploit) that further democratize the study of adversarial machine learning and facilitate real-time experimentation for deep learning practitioners.

Finally, we also present how scalable interactive visualization can be used to amplify people’s ability to understand and interact with large-scale data and complex models. We sample from projects where interactive visualization has provided key leaps of insight, from increased model interpretability (Gamut with Microsoft Research), to model explorability with models trained on millions of instances (ActiVis deployed with Facebook), increased usability for non-experts about state-of-the-art AI (GAN Lab open-sourced with Google Brain; went viral!), and our latest work Summit, an interactive system that scalably summarizes and visualizes what features a deep learning model has learned and how those features interact to make predictions. We conclude by highlighting the next visual analytics research frontiers in AI.   

Live

Many complexity-rich dynamical systems necessitate the use of a multiparameter probabilistic model to capture the observed system behavior succinctly. Unfortunately, multiparameter probabilistic models of large systems suffer from the curse of dimensionality. To alleviate this problem, we recently proposed a manifold embedding approach that borrows a concept from special relativity: the intensive symmetrized Kullback Liebler (isKL) embedding [1]. This approach generates an analytically tractable embedding for model predictions in Minkowski space, for most common probability distributions and statistical models. In principle, this technique not only offers a low dimensional representation of high dimensional data, but it also allows one to uncover hidden exponential families that describe experiments or simulations. In this talk, we will showcase how this technique can be combined with a probabilistic neural network to study cartilage tissue and bird song data.

[1] Teoh, et al. Phys. Rev. Research 2.3 (2020): 033221   

Live

Machine learning approaches including neural networks (NN) and Gaussian process regression (GPR) are finding widepread use to recover functional dependencies from multidimensional data. As powerful as these approaches are, they may fail when data density is low, which is always the case in highly-dimensional cases. Some methods like GPR also cannot easily work with large datasets. Using modified high dimensional model representation (HDMR) to represent a multivariate function with machine-learned lower-dimensional terms allows recovering functions from very sparse data, down to ~2 data per dimension. Sub-dimensional component functions are easier to fit and to use. Specifically here we present a HDMR-GPR combination where the use of GPR to represent component functions allows nonparametric (unbiased) representation and the possibility to work only with functions of desired dimensionality, obviating the need to build an expansion over orders of coupling. All component functions are determined from a single set of samples. We test the method by fitting potential energy surfaces of polyatomic molecules as well as by computing vibrational spectra.   

Live

A key challenge in searches for resonant new physics is that classifiers trained to enhance potential signals must not induce localized structures. Such structures could result in a false signal when the background is estimated from data using sideband methods. A variety of techniques have been developed to construct classifiers that are independent from the resonant feature (often a mass). Such strategies are sufficient to avoid localized structures, but are not necessary. We develop a new set of tools using a novel moment loss function (Moment Decomposition or MoDe) which relax the assumption of independence without creating structures in the background. By allowing classifiers to be more flexible, we enhance the sensitivity to new physics without compromising the fidelity of the background estimation.   

Live

Data visualization is critical to understanding data in physics. Visualization is important in the early stages of understanding a measurement or a model and, later, in communicating it clearly. Physicists need to produce visualizations that fit their specific problem, often including multiple axes or novel plot types. They need tools that make common tasks like scatter and line plots easy and complex tasks like compound figures possible.

Matplotlib is the foundational data visualization library for the Scientific Python Ecosystem. It is used by working physicists, from high-profile projects like LIGO and the Event Horizon Telescope. to the day-to-day work of students, for interactive exploratory visualization and publication-ready figures.

This talk will highlight some of the key features of the library, focusing on examples of interactive multi-scale visualizations and tuning figures in preparation for publication. I will also briefly discuss our future plans to evolve the library to meet the future needs of our users.   

Live

The deep extention of the more popular restricted Boltzmann machine (RBM), known as deep Boltzmann machines (DBMs), are expressive machine learning models which can serve as compact representations of complex probability distributions. However, jointly training DBMs in the unsupervised setting has proven to be a formidable task. A recent technique we have proposed [1], called mode-assisted training, has shown success in improving the unsupervised training of RBMs. Here we show that indeed the performance gains of mode-assisted training translate to the DBM as well, compared to the baseline approach based exclusively on Gibbs sampling. Furthermore, we find evidence that DBMs trained with the mode-assisted algorithm can represent the same data set with fewer total weights compared to RBMs. We perform a comparison on small synthetic data sets where exact log-likelihoods are computed, as well as the popular MNIST standard.

[1] H. Manukian, et al. Comm. Phys. 3, 150 (2020)   

Live

The atomistic structure of complex nanoscale structures such as the grain boundaries or the nanoclusters is difficult to determine even using state-of-the-art experimental characterization techniques. So, the use of theoretical characterization methods is beneficial to address this problem. In this talk, we will discuss the modular FANTASTX (Fully Automated Nanoscale To Atomistic Structures from Theory and eXperiments) toolkit which explores the potential energy landscape for low energy structures that also match with experimental data. Experimental data types include scanning tunneling microscopy (STM), transmission electron microscopy (TEM), and x-ray pair distribution function (PDF), among others. Energetic information obtained from density functional theory or empirical potential calculations is used in conjunction with match to experimental data to guide the search. We will discuss the use of artificial intelligence/machine learning to accelerate the search. We will describe how FANTASTX is used to determine the atomistic structures of interfaces observed in STEM, nanoclusters observed with PDF, and surfaces observed with STM.   

Live

Diffraction-based experimental techniques like X-ray Reflectivity (XRR) determine the distribution of electrons at the surface of a crystalline solid but inverting this data to obtain the atomic structure of the surface is a challenge. To overcome this obstacle, we develop Python software which optimizes the surface structure by utilizing energetic information from DFT alongside data from multiple experimental measurements under different conditions. Our work leverages Python object orientation and scientific libraries to create modular and flexible software with access to powerful optimization techniques. Using the SrTiO3 (001) surface as an example, we determine interfacial structure using known low energy surface terminations from DFT and experimental measurements from X-rays which are resonant and non-resonant with the Sr K-edge. Using this joint experimental-theoretical approach to investigate the interfacial structure-property relationship provides insights which could increase the performance of diverse technologies such as energy storage and conversion and semiconductor fabrication.   

Live

Predictive Science involves observational data collection, developing testable hypotheses, & making predictions. The scientific theory developed can be cast into a mathematical objective function that goes from various steps of model calibration, validation, & prediction of the Quantity of interest.
For decades, the Markov Chain Monte Carlo algorithms, especially the Metropolis-Hastings algorithm, are widely used for stochastic optimization, sampling, & integration of mathematical objective functions, in the context of Machine Learning, Bayesian inverse problems, & parameter estimation. Here, we present the ParaMonte software, a suite of parallel Monte Carlo optimization, sampling, & integration algorithms for Bayesian inference problems.
The primary goal of the ParaMonte library is to streamline scientific inference by full automation & by providing runtime dynamic directions to the user. It also offers fully-deterministic reproducible restart functionality of all simulations & a unified API accessible from several major scientific & Data Science programming languages, including C, C++, Fortran, MATLAB, Python, & R. Comprehensive automated post-processing tools integrated with the ParaMonte library also enable seamless analysis & visualization of the simulation results.