Advancing the CANWET watershed model and decision support system by utilizing high performance parallel computing functionality
Watershed modeling is widely used to better understand processes and help inform planning and watershed management decisions. Examples include identifying impacts associated with land use change; investigating outcomes of infrastructure development, predicting effects of climate change. The proposed project will see the evolution of a desktop based watershed modeling and decision support system to a web based tool that will allow greater access by decision makers and stakeholders. By this means we will advance the idea of evaluating cumulative effects in the watershed decision making process rather than the current practice of assessing proposed changes in isolation.
The proposed software evolution will take advantage of high performance computing by porting existing code to a higher performing language and restructuring to operate using parallel or multi-core processing. The result is expected to be a dramatic reduction in simulation run times. Reduced run times will facilitate the use of automatic calibration routines used to conduct model setup, reducing costs. It will also enable rapid response if the simulation were to be re-run by a request through the web-based user interface. The designed web-based tool will be used by decision and policy makers in the watersheds that contribute to Lake Erie to understand the sources of pollution especially phosphorus which is a major contributor to Lake Erie eutrophication problems and develop policies in supporting a wide variety of watershed planning and ultimately help achieve the Federal and Ontario government commitments to reduce 40% phosphorus entering Lake Erie by 2025.
Automated Assessment of Forest Fire Hazards
The proposed project is to develop a Deep Reinforcement Learning (DRL) model capable of using Light Detection and Ranging (LiDAR) data to automatically analyze whether there is a risk of forest fire due to vegetation management around electric equipment. The model automatically classifies points in a LiDAR scan to determine whether there is direct contact between vegetation and electric equipment and outputs control actions for an Unmanned Aerial Vehicle (UAV) that can autonomously inspect infrastructure as needed. Our research will focus on the solution of two different but complementary machine learning problems that the DRL model will have to solve to allow for a UAV to autonomously acquire data with sufficient quality for fire risk assessment. Firstly, there is the problem of perception, i.e., to analyze raw LiDAR data and deduce classifications of individual objects and their shape and orientation. Secondly, there is the problem of control, i.e., to use the perceived state of the environment as determined by the first stage to adequately navigate through the environment until a complete scan has been obtained of the region that we are conducting a fire risk assessment for. Solving these two problems in a harmonious way that allows for an efficient, accurate and a complete scan of the environment is an open research problem in the domain of DRL that we aim to address.
Generative models for autonomous, video-based DJ systems
RaveDJ is all about enabling people to create music with AI. Over the past few years our team has worked tirelessly to make RaveDJ the world’s first fully autonomous DJ system. A system that can seamlessly mix together a set of songs to produce the optimal mix to listen to as well as unique combinations or mashups of two songs into one. RaveDJ is currently producing around 10-50 TB per day of mashups and mixes for people all around the world. RaveDJ is hosted on our website www.rave.dj as well as our app Rave (get.rave.io/download) which has millions of users from around the world. SOSCIP’s platform will enable us to investigate ways of using bleeding edge machine learning techniques to navigate around existing inefficiencies in our existing system, as well as to take a leap beyond its current level of performance with new research advances. SOSCIP’s platform is exceptional for helping us achieve these goals as it can support training of deep neural networks on very large datasets of music videos. This is a requirement for analyzing and generating thousands of high-resolution videos at record speed. The platform that SOSCIP provides is perfect for running the experiments that are required to produce models that can perform these tasks efficiently as well as addressing new and novel problems in using AI for creative applications. The collaboration is between Rave, Prof. Graham Taylor and Prof. Stefan Kramer at University of Guelph.
High-throughput transcriptomic and eQTL analyses of silicon-induced resistance against Fusarium head blight on wheat
Wheat is an essential part of the agriculture and agri-food industry in Ontario with an average annual production of 2.5 Million tons. High yield and quality of wheat is seriously threatened by biotic and abiotic factors, among which Fusarium head blight (FHB, caused by the fungus Fusarium graminearum) has historically been most damaging. The mycotoxin produced by this fungus is harmful for human health and livestock feed and productivity. Conventional chemical fungicides are commonly used as an important mean to control FHB; however, they also pose a serious risk to the health of humankind and the environment. One of the most promising non-hazardous, eco-friendly methods to control different plant pathogens is the use of silicon, which induces resistance.
Backed by SOSCIP high-performance computing resources, and using novel bioinformatic approaches, this study aims at conducting high-throughput genome-wide transcriptomic analysis of silicon-induced resistance against FHB, and possibly other important diseases, in wheat. Such study will lead us to:
1. understanding the complex underlying induced defense mechanisms
2. identifying gene modules and regulatory elements that control such mechanisms, and finally and more importantly
3. providing valuable building blocks and framework for future breeding programs which will be focused on the development of novel disease-resistant wheat cultivars in Ontario.
Real-time flood forecasts using parallel cloud computing and intelligent algorithms
Development of flood forecast early warning systems is critical as storm and spring melt events intensify due to climate change. Predicting these events with precision and adequate warning time presents a significant technological and computational challenge for managers of water resource systems.
Effective flood forecasting and warning systems combine several complex modelling tools. Climate forecast, hydrologic and hydraulic models are used together to forecast flows, water levels and flood inundation extents. Such models are computationally expensive, and each has associated uncertainties that must be quantified.
This project addresses these issues by coupling models with Machine Learning algorithms and high-performance computing infrastructures. The approach will enable the existing flood forecasting platform (known as ISWMS) to better i) quantify uncertainties associated with flood forecasts and ii) minimize run-time of the computationally expensive models embedded within the forecasting system.
Sparse Representations for Embodied AI
Understanding scenes representing real-world environments is a challenging problem at the intersection of Computer Vision research and Deep Learning and a necessary pre-requisite for Embodied AI. Embodied AI is an emerging field within Machine Learning that focuses on the challenges that need to be addressed for the successful deployment of edge devices such as drones and robots. In this setting estimating the semantics of an environment plays an essential role in addition to how it can be efficiently navigated to solve a variety of tasks that can involve other agents as well. The Embodiment Hypothesis states that intelligence emerges from the interaction of an agent and its perception of the environment it is embodied within. This project establishes an empirical study to validate this hypothesis for Deep Reinforcement Learning (DRL) agents trained on environments derived from simulations as well as real-world data. We use DRL as a methodology to model a Partially Observable Markov Decision Process (POMDP) that describes optimal processes for navigation-perception problems. The embodiment hypothesis implies that an end-to-end understanding of this problem should outperform conventional methods for training computer vision models. To address the limitations of existing DRL algorithms on this task, we propose a novel family of networks that learn to solve embodied tasks from sparse representations of the perceived data. Our aim is to enable a new paradigm for efficient production systems for drones to navigate complex environments and visually monitor infrastructures such as for autonomous fire risk assessment and asset monitoring.