Paper No. 5
Presentation Time: 2:50 PM


WONG, Cindy M.1, HARRISON, Adam P.1, RANAWEERA, Kamal2 and JOSEPH, Dileepan1, (1)Electrical and Computer Engineering, University of Alberta, ECE Research Facility, Edmonton, AB T6G 2V4, Canada, (2)Arts Resource Centre, University of Alberta, Arts & Convocation Hall, Edmonton, AB T6G 2E6, Canada,

Artificial Intelligence (AI) describes a computing system able to perceive, analyze, and respond in an appropriate manner to maximize success in a given context. Significant research focuses on AI problems, covering a wide range of contexts. To help solve the AI-complete problem of image understanding, contexts of sufficient importance must be chosen because human input is essential given the state of the art today. Microfossil identification is one such context.

Microfossils, in particular marine microfossils, are important because of the rich information they have captured over millennia. For example, geoscientists collect and identify fossilized shells of foraminifera to map present day hydrocarbon accumulations, through biostratigraphy, and to infer prehistoric environmental conditions, through isotope geochemistry. For best results, specimen genus and species are identified, a labor intensive process normally done manually. Vast numbers of microfossils have been collected by ocean drilling programs, but only a tiny fraction have been identified.

Automation of microfossil identification has been attempted since the late 1980s, initially with rule-based systems, such as Fossil and VIDES. These systems helped knowledgeable users refine a list of possible taxonomic identities interactively. In general, rule-based approaches required users to examine each specimen under a microscope and identify features manually. Later systems based on Artificial Neural Networks (ANNs), such as CLASSIC, COGNIS Light, and SYRACO, were more effective at reducing labour. However, ANN-based approaches were also inadequate for reasons ranging from issues with performance to issues with scalability of performance.

Thus, we investigate human-based computation for microfossil identification, engineering a system by iterative and incremental development. With this approach, computers do as much image understanding as they can and outsource the rest, over the Internet, to humans. Less crowdsourcing will be needed as the project evolves. We engineer technology, such as virtual reflected-light microscopy and dynamic hierarchical identification, to integrate computer and human intelligence efficiently. In general, human-based computation is an increasingly popular approach to tackle AI-complete problems.

  • GSA 2012.pptx (7.9 MB)