Previous seminars
Chrissy Fell and Ben Baer
Speaker: Dr Chrissy Fell, University of St Andrews
Title: Examples of deep learning applied to medical and ecological images.
Abstract: In this talk I will discuss three projects I have been working on applying deep learning to images. The first project is creating an automated classifier for images of endometrial and cervical biopsies that allows prioritisation of pathology workloads. The second project I will talk about is automatically detecting animals in aerial images. Finally I will explain my recent work on using brain MRI scans from the UK Biobank to classify if someone is at high or low genetic risk of mental health condition.
Speaker: Dr Ben Baur, University of St Andrews
Title: Some problems I’m working on
Abstract: The talk has three parts in which an overview of an estimation framework I now commonly use is sandwiched by some problems I’m working on. In the first part, a problem involving a barely identified discrete parameter in a highly structured model is presented. After, the failure of various Bayes estimators with non-informative priors is briefly described. In the second part, some aspects of data coarsening and semi- and non-parametric efficiency theory are explained alongside examples from causal inference and survival analysis. In the third part, several ongoing projects involving coarsening or efficiency theory are presented with a frame or two per project. The audience is encouraged to frequently stop me with comments and questions.
Broad-scale estimates of detection probability in North American landbirds, with implications for data integration
Speaker: Brandon Edwards, Carleton University, Canada
Abstract: Since the 1970s, North America has lost approximately 2.9 billion birds, despite decades of targeted conservation efforts. In a world where biodiversity monitoring data is increasing year after year, but biodiversity itself continues to decrease, methods and expertise from the world of Big Data are now needed to properly synthesize the thousands of bird monitoring datasets into one cohesive story. Using hundreds of open datasets across North America, the NA-POPS project has derived detection probabilities for nearly 75% of North America’s landbirds. Here, I will present the results of this broad-scale effort to estimate accurate detection probabilities for North American landbirds. Additionally, I will present some preliminary results from two additional projects that have stemmed off this effort: 1) a modelling exercise to predict detection probabilities for rare or undersampled birds using a hierarchical Bayesian model, and 2) a framework to estimate distance to singing birds (i.e., detection distance) using autonomous recording units. Finally, I will talk about my proposed project that I am undertaking at University of St Andrews in conjunction with CREEM to apply these detection probabilities as statistical offsets in an integrated trend modelling framework, to improve data coverage and (hopefully) trend estimates of the North American Breeding Bird Survey.
Self-exciting point process models
Speaker: Charlotte Jones-Todd, University of Auckland, New Zealand
Abstract: Modelling spatial and temporal patterns in ecology is imperative to understand the complex processes inherent in ecological phenomena. Log-Gaussian Cox processes are a popular choice amongst ecologists, used to describe the spatiotemporal distribution of point-referenced data. In addition, self-exciting point pattern models are becoming increasingly popular to infer the contagious nature of events (e.g., animal sightings, cue rates). In this talk, I will present extensions to these well-known models that include spatiotemporal self-excitation and joint likelihood models. Such models are, often, better equipped to capture the complex mechanisms inherent in many ecological and environmental data.
Dealing with the badness of goodness-of-fit
Speaker: Rishika Chopara, University of Auckland, New Zealand
Abstract:
Goodness-of-fit (GOF) testing is vital to statistical analysis, as it allows us to validate the reliability of any statistical inference we make.
In many statistical models, the deviance is used to assess GOF by comparing it against a Chi-squared distribution. However, in some situations (e.g. when dealing with sparse counts) the deviance does not have a Chi-squared distribution, even approximately, yielding such tests unusable. In principle, the true distribution for the deviance is computable, however in practice it is often intractable. We show that generally, we can accurately approximate the true underlying distribution of the deviance using a Gamma distribution. Using this approximation, we enhance the usability and power of GOF tests while retaining the familiarity and convenience of the deviance statistic.
Using a range of capture-recapture models for illustration, we show how our method can be used to accurately approximate the distribution of the deviance when dealing with various levels of data sparsity. With this approach, we aim to provide an accessible and effective GOF testing framework for complex modelling scenarios such as spatial capture-recapture.
Postponed – Broad-scale estimates of detection probability in North American landbirds, with implications for data integration
Speaker: Brandon Edwards, Carleton University, Canada
Abstract:
Since the 1970s, North America has lost approximately 2.9 billion birds, despite decades of targeted conservation efforts. In a world where biodiversity monitoring data is increasing year after year, but biodiversity itself continues to decrease, methods and expertise from the world of Big Data are now needed to properly synthesize the thousands of bird monitoring datasets into one cohesive story. Using hundreds of open datasets across North America, the NA-POPS project has derived detection probabilities for nearly 75% of North America’s landbirds. Here, I will present the results of this broad-scale effort to estimate accurate detection probabilities for North American landbirds. Additionally, I will present some preliminary results from two additional projects that have stemmed off this effort: 1) a modelling exercise to predict detection probabilities for rare or undersampled birds using a hierarchical Bayesian model, and 2) a framework to estimate distance to singing birds (i.e., detection distance) using autonomous recording units. Finally, I will talk about my proposed project that I am undertaking at University of St Andrews in conjunction with the Centre for Research into Ecological and Environmental Modelling to apply these detection probabilities as statistical offsets in an integrated trend modelling framework, to improve data coverage and (hopefully) trend estimates of the North American Breeding Bird Survey.
Keeping it discrete: challenges with population modeling of continuous processes
Speaker: Dan Linden, Northeast Fisheries Science Center (NOAA)
Using distance sampling to understand how sea-going citizen scientists detect marine birds
Speaker: Michael Schrimpf, Cornell University
Abstract: Ecotourists are increasingly crossing the open ocean on cruises, collecting citizen science data on seabirds as they do. These data can be useful for monitoring species distributions, but only if factors affecting the detection of birds from such ships are better understood. We partnered with the Antarctic expedition cruise industry to calibrate citizen science data from the eBird project (https://ebird.org/) by conducting concurrent distance sampling surveys aboard tour vessels. Our surveys measured factors that likely add bias to data collected by eBird volunteers, such as the observer’s vantage point and the tendency of birds to follow vessels. However, we also faced the challenge of distance sampling with fast-moving birds, and we are currently exploring simulations to help correct for this violation of distance sampling assumptions. Although this project is still ongoing, our preliminary results include several species-specific seasonal distribution patterns not previously described that are likely related to post-breeding dispersal. Our ultimate goal is for the data-rich, underused collection of at-sea eBird observations to aid in seabird monitoring and conservation efforts.
Integrating individual movement processes into population models
Speaker: Beth Gardner, University of Washington
Abstract: Understanding of the ecological processes that drive populations requires knowledge not only of demographic rates (e.g., abundance, survival, reproduction) but also of how animals use space through resource selection and movement. For example, individual movement is a key driver of population dynamics for recolonizing species. To incorporate both demographic and movement processes requires a flexible framework, one that can integrate data from multiple streams, such as an integrated population model. In this presentation, I’ll demonstrate approaches to incorporating individual movement into spatial capture recapture models and integrated population models. I’ll present a case study on recolonizing grey wolves (Canis lupus) in Washington, USA. We developed a model for grey wolf recolonization that has two main components: [1] an age- and state-structured population model that governs the population state process, and [2] an individual-based spatial model describing the dispersal of individuals and colonization of sites. We also considered different movement models to incorporate uncertainty in how dispersing wolves select new territories, an essential but unobservable process. All forms of the model resulted in showing gray wolves have a >99% probability of colonizing all of Washington State’s recovery regions by 2030. This case study, along with the other models presented, highlights how incorporating movement processes into population models and propagating uncertainty throughout the model can help us to address ecological questions related to habitat use, selection, and population dynamics.
Seminar – Fast fitting of NODEs by Bayesian neural gradient matching.
Speaker: Willem Bonnaffeé, Nuffield Department of Surgical Sciences, University of Oxford
Abstract: Neural ordinary differential equations (NODEs) allow us to nonparametrically infer ecological interactions from time series data. NODEs however are slow to fit. We developed a fitting method, Bayesian neural gradient matching (BNGM), which reduces NODEs fitting times to only a few seconds. We test the approach in an artificial tri-trophic system where true ecological interactions are known, as well as in an experimentally replicated time series of an algae, flagellate, and rotifer, and in the multi-species aquatic community of the Maizuru bay in Japan. We show that BNGM is faster and more accurate than standard NODEs, parametric ODEs, and convergent cross-mapping (CCM). Our results suggest that only main interactions are consistent across the replicated time series, and that dynamics of the Maizuru bay community are largely driven by a single species. Overall, this work shows that while NODEs alleviate the need for a mechanistic understanding of interactions, BNGM alleviates the heavy computational cost, thus greatly improving the usability of the approach.
Development of a new control rule for managing anthropogenic removals of protected, endangered or protected species in marine ecosystems
Speaker: Matthieu Authier, Pelagis Observatory, France
Title: Development of a new control rule for managing anthropogenic removals of protected, endangered or protected species in marine ecosystems
Abstract: By-catch, the unintentional catch of other species during fishing operations targeting commercial species, is a major threat to many marine Protected, Endangered or Threatened Species (PETS) including cetaceans. The aim of this work was to propose a new rule for managing anthropogenic removals of PETS, compared to existing rules. We developed a stochastic SPM model (Surplus Production Model) which includes abundance and removal processes, linked to each other under the assumption that variations in removals reflect variations in abundance. By-catch and abundance are linked by a removal rate (ϕ). The three other parameters of the stochastic SPM are: the intrinsic growth (r), the carrying capacity (K) and an environmental stochasticity (σ). A simulation study was carried out under the SPM to investigate (1) the sensitivity of population extinction to {ϕ, r, σ}, and (2) parameter identifiability from typical data that can be collected on PETS.
We then consider several control ’harvest’ rules to manage PETS by-catch, included two new ones from this SPM model. The two new rules hinge on the estimation of a stationary removal rate (ϕ). We compared these new rules to other existing control rules (e.g. Potential Biological Removal or a fixed percentage rule) in three scenarios: (i) a base scenario whereby unbiased but noisy data are available, (ii) scenario whereby abundance estimates are overestimated and (iii) scenario whereby abundance estimates are underestimated. The different rules were tested on a simulated set of data with life-history parameters close to a small-sized cetacean species, the Harbour Porpoise (Phocoena phocoena) and in a MSE (Management Strategy Evaluation) framework. The effectiveness of the rules were assessed by looking at performance metrics, such as time to reach the conservation objective, the removal limits obtained with the rules or temporal autocorrelation in removal limits. One new rule relying on the SPM model and the trend in abundance was robust against biases in data and displayed greater alignment with EU conservation objectives. This new rule is promising and could be used in the management of marine mammals by-catch, such as the common dolphin (Delphinus delphis) in the Bay of Biscay.