Invited Speakers

Confirmed invited speakers:

1. Gustau Camps-Valls, Universitat de Valencia 

Unsupervised Deep Feature Learning with Sparse Codes and Gaussianization

In this talk I'll review our latest works on unsupervised feature extraction with hierarchical (deep) representations. Two different pathways will be taken: a phisiologically /meaningful/ representation with deep convolutional neural networks where both population and lifetime sparsity is imposed, and a /meaningless/ (projection pursuit like) cascade representation where the goal is to transform the data into a multivariate Gaussian. I will discuss about the advantages and shortcomings, and illustrate their performance in synthetic and real problems of image segmentation, data classification, synthesis, and information estimation.

2. Julien Emile-Geay, University of Southern California

Paleoclimate informatics: enabling knowledge discovery about past climates

Climate exhibits scaling behavior, which means that fluctuations increase in amplitude with the timescale. Some of the most scientifically and societally-relevant fluctuations occurred before the short instrumental record beginning ca 1850, making their study impossible with such observations. Paleoclimate (pre-instrumental) observations are thus a critical window into this behavior, but pose a unique set of challenges to the analyst: they are indirect, typically sparse and noisy, and come from incredibly diverse archives, impeding a one-size-fits-all approach.


Until recently, integrating all these data sources into a coherent picture was impossible. In this talk, I will review recent progress in paleoclimate informatics, including paleo data semantics, data assimilation and graph theory — all the product of joint work with Y. Gil, L. Bradley, N. McKay, D. Guillot and the Last Millennium Reanalysis collective. I will argue that these advances now enable thoughtfully applied machine learning algorithms to play a useful role in knowledge discovery, with the potential to advance the study of past climates.  

3. Lucas Joppa, Microsoft, AI for Earth

AI for Earth

The speed and scale at which climate systems are changing, and the enormity of the human impact of those changes, requires a commensurate response in how society monitors, models, and manages climate systems. A key component to that response will emerge from the fundamentals of AI – transforming how we collect data, convert those data into actionable information, and communicate that information across the world. By training increasingly sophisticated algorithms with this unprecedented collection of data on dedicated computational infrastructure, we can combine human and computer intelligence in a way that will allow us to make increasingly informed and optimal choices about today – and tomorrow.

4. Eric Maloney, Colorado State University

Critical challenges in the simulation of tropical clouds and climate

Realistic simulations of current climate and projections of future climate change require that the treatment of clouds and moist convection in models be realistic. However, many climate models exhibit substantial biases in various aspects of the climate system that depend on realistic cloud parameterizations, which affects confidence in their ability to simulate future climate changes. Indeed, cloud feedbacks are one of the greatest sources of uncertainty in projections of future climate. This talk will highlight some tropical biases that are related to treatment of cloud parameterizations including in the simulation of the Madden-Julian oscillation, low clouds in colder tropical regions, and monsoon systems. The rectification of these biases onto aspects of climate such as the global mean energy budget will be discussed, especially as they affect future climate projections. Community efforts to diagnose these biases by entraining observations into process-oriented diagnostic frameworks will be highlighted, including a recent effort by the NOAA Model Diagnostics Task Force. New modeling approaches to parameterization of clouds in climate models will be discussed that provide great promise to mitigate tropical biases, and the talk will close by discussing some initial attempts to use machine learning to parameterize cloud processes in models.

5. Christopher Wikle, University of Missouri

Using parsimonious “deep” models for efficient implementation of multiscale spatio-temporal statistical models applied to long-lead forecasting

Spatio-temporal data are ubiquitous in engineering and the sciences, and their study is important for understanding and predicting a wide variety of processes. One of the chief difficulties in modeling spatial processes that change with time is the complexity of the dependence structures that must describe how such a process varies, and the presence of high-dimensional complex datasets and large prediction domains. It is particularly challenging to specify parameterizations for nonlinear dynamical spatio-temporal models that are simultaneously useful scientifically and efficient computationally. One potential parsimonious solution to this problem is a method from the dynamical systems and engineering literature referred to as an echo state network (ESN). ESN models use so-called reservoir computing to efficiently compute recurrent neural network (RNN) forecasts. Moreover, so-called ``deep" models have recently been shown to be successful at predicting high-dimensional complex nonlinear processes, particularly those with multiple spatial and temporal scales of variability (such as we often find in spatio-temporal geophysical data). Here we introduce a deep ensemble ESN (D-EESN) model in a hierarchical Bayesian framework that naturally accommodates non-Gaussian data types and multiple levels of uncertainties. The methodology is first applied to a data set simulated from a novel non-Gaussian multiscale Lorenz-96 dynamical system simulation model and then to a long-lead United States (U.S.) soil moisture forecasting application.

6.  Qi (Rose) Yu, Northeastern University

Deep Learning for Large-Scale Spatiotemporal Data

In many real-world applications, such as climate science, transportation and physics, machine learning is applied to large-scale spatiotemporal data. Such data is often nonlinear, high-dimensional, and demonstrates complex spatial and temporal correlations. Deep learning provides a powerful framework for feature extraction, but existing deep learning models are still insufficient to handle the challenges posed by spatiotemporal data.

In this talk, I will show how to design deep learning models to learn from large-scale spatiotemporal data. In particular, I will present our recent results on 1) High-Order Tensor RNNs for modelling nonlinear dynamics, and 2) Diffusion Convolutional RNNs for modelling spatiotemporal patterns, applied to real-world climate and traffic data. I will also discuss the opportunities and challenges of applying deep learning to large-scale spatiotemporal data.