TIES 2018 Short Course on Deep Learning for Environmental Sciences
Dr. Nathaniel K. Newlands, CANADA
Modern machine learning and artificial intelligence procedures provide fast and automatic learning of hidden dependencies and structures. It offers many advantages for operational predictive processes and systems by providing higher automation/speed, accuracy, scalability, and strategic foresight. The benefits of deep learning is being realized across a wide variety of real-world applications from speech and image recognition, smart application communications, communications, medicine/drug discovery, manufacturing, equipment failure and flaw detection, transportation to threat/fraud/risk management, and environmental science. Challenges with modeling the dynamics of complex ecosystems also brings to the fore statistical problems of analyzing massive, multi-resolution, multi-source data with a nonstationary space-time structure – where deep learning may help provide an integrated solution. In particular, reinforcement algorithms that incorporate deep learning help to predict within the so-called ‘delayed return’ environment where it can be difficult to understand which action leads to which outcome over many time steps. Here, deep learning may enable near-real-time forecasting and strategic foresight in the more ambiguous ‘real-world’ environment, whereby one chooses from an arbitrary number of possible actions to achieving goals.
This short course is relevant to anyone with an interest in Environmental Science and Sustainability, Data Science, and Predictive Analytics. Previous training and experience in predictive modeling, machine-learning and R Statistical Language are an asset. The course will being with an introduction to artificial intelligence, machine-learning and neural networks. An overview of network model architecture, learning structures (i.e., multilayer perceptrons (MLPs), convolutional, recurrent, deep) and supervised/unsupervised optimization and learning algorithms will be discussed. State-of-the-art toolsets and code libraries for tackling deep learning tasks will be surveyed and applied (e.g., h20, Deepnet, Keras, DeepLearning4J, TensorFlow). A set of environmental science examples will be used to illustrate core concepts, features, and benefits. These case examples will include: predicting wine quality, plant phenomics, crop yield, extreme events, and air pollution.
Additional details and information, including relevant R libraries required and example codes will be provided to participants via email (i.e. a dropbox shared folder) in advance of the workshop. Labtops with an internet connection will be provided for participants. Participants should install the latest version of R Statistical Software and R Studio Desktop.
R version 3.4.3 (latest)
R Studio Desktop (free version available)