Seminar Schedule in Google Calendar
Room 4192, Earth Sciences Bldg (2207 Main Mall)
Tue 17th May 2016
Anomaly Detection in Time Series Datasets of Internet of Things Devices
This presentation will describe advances in identifying anomalies in datasets of internet of things (IoT) devices found in commercial buildings. Leveraging properties specific to machine-to-machine communications, data models and algorithms can be constructed from the time series datasets to identify security threats (i.e. cyber attacks) and failing devices (predictive maintenance). Devices modeled include security cameras, lighting and heating control, and various sensors and actuators.
Big data techniques can also be applied to the communication datasets, including various forms of clustering and fuzzy logic/fuzzy inferencing to identify deviations from real-world normal operations. In such situations, modularity is key to proper implementation in order to allow for proper data processing, algorithm training and algorithm testing.
Results of current research and future roadmap will be discussed.
About Optigo Networks
Optigo Networks is shaping the future of the commercial Internet of Things (IoT) by redefining how smart buildings are connected and operated. By applying visualization and anomaly detection to the building system, Optigo allows the IoT to scale, driving down the cost to maintain and operate the technologies that make buildings comfortable and efficient.
With its award-winning software, Optigo Networks allows building operators to quickly identify faults and security threats in the building system, cutting troubleshooting time down from hours to minutes. Builtin analytics rein in the building IoT, reducing OpEx and maintenance costs with tools and reports to visualize the health and security of the building network.
Room 4192, Earth Sciences Building (2207 Main Mall)
Tue 10th May 2016
Gal Av-Gay, MSc Student, UBC Statistics
Optimizing the Metropolis Hastings algorithm for Gaussian Processes
11:00am - 11:30am
The Metropolis-Hastings algorithm is often used to obtain Markov Chian Monte Carlo samples from highly complex posterior distributions, such as those involved in full inference of hyper-parameters in Gaussian Process models. Here we compare one new and three existing implementations of the Metropolis-Hastings algorithm in the context of sampling from the posterior distributions of hyper-parameters in a Gaussian Process. The implementations are compared in terms of their initialization biases and convergence rates, as well as in terms of their performance on higher dimensional data. Our experiments involve sampling from GP posterior distributions using the four different implementations and comparing the quality of these samples. A discrepancy measure is devised based on the Kolmogorov-Smirnov test to measure the convergence rate of each algorithm. Issues in generating MCMC samples for high dimensional data are discussed and an optimal approach to sampling is proposed called the Laplace approach. All of the algorithms in this paper are implemented in an R package called ‘gpMCMC’.
AERL 120 (2202 Main Mall)
Wed 4th May 2016
Tom Carruthers, Research Associate; Quantitative Modeling Group (Institute for the Oceans and Fisheries, UBC)
New methods in fisheries modelling and statistics
In a three part review of my latest research I describe new developments and findings in (1) landscape-scale social ecological systems modelling for a BC recreational fishery, (2) statistically rigorous approaches for accounting for uncertainty from data processing in population analyses and (3) using simulation to establish robust management for Atlantic bluefin tuna and data-limited global fisheries.