Crowdsourcing: combining human and machine elements

Julian Jarrett, Iman Saleh, M. Brian Blake, Rohan Malcolm, Sean Thorpe, and Tyrone Grandison presented their paper entitled “Combining Human and Machine Computing Elements for Analysis via Crowdsourcing” at CollaborateCom 2014.

The work investigates three research questions:

  1. When human and machine elements are capable of performing the same task, is there a general model that can define and evaluate their respective performance outcomes simultaneously?
  2. Can experimentation in a specific domain, such as face recognition, uncover the most appropriate shared evaluative attributes that have cross-domain applicability?
  3. Can the specific performance variations in real-life experimentation enhance our overall understanding and ultimately lead to a more generalized elastic model?

To address these questions the researchers device an elasticity framework consisting of elasticity manager, resource manager, and solution formulation. The framework manages the provisioning of machine computing elements (MCEs), human computing elements (HCEs), or both referred to as Elastic Computing Elements (ECEs). The algorithms use the elasticity model to ascertain the complexity of the task and to enforce constraints on its successful completion.

What is the main idea of the framework? The main idea is to extract the elasticity attributes of a certain task and use these attributes to orchestrate the use of HCEs and MCEs. The researchers affirm that this study differs from other projects as it considers the situations, where the task for humans and for machine elements are exactly the same.

A face recognition task  is used to evaluate the framework, in which testing images of the same individuals pass through the elasticity system for identification. Face recognition tasks are originally sent to the MCE component and then forwarded to HCE components via mobile crowdsourcing app.

The results are clear: using the proposed elasticity framework, the probability of positive identification increases significantly. Moreover, the researchers illustrate the connection between task’s TCI and the performance index of a computing element and show that ECEs render the most optimal results. Based on this, the authors propose the Maximum Performance Algorithm, which in most cases should generate the maximum performance index and the Weighted Metric Performance Index Algorithm which should provide the most optimal results.

The full paper can be found here.