Crowdsourcing utilizes the input of a crowd of online users
to collaboratively solve problems. To advance this emerging
technology, researchers at the University of Miami are
developing a computing model that uses crowdsourcing to combine
and optimize human efforts and machine computing elements.
The new model can be used to efficiently perform the complex
tasks of face recognition — a method used in law
enforcement. It’s a new approach to using social networks as a
formal part of the criminal investigation process, explained
Brian Blake, vice provost for Academic Affairs and dean of the
Graduate School at the University of Miami (UM). He is also
principal investigator of the project.
“The breadth of the internet and popularity of smartphones
have facilitated the onset of online crowdsourcing platforms,”
Blake said. “Our project attempts to leverage the power of the
crowd to solve complex problems, on demand.”
Preliminary findings showed an average certainty of 14.13
percent for machine computing elements (MCE) to identify
individuals in pictures. Combining the efforts of both MCE and
human computing elements (HCE) in performing the same task,
there was an average increase of 55 percent reflecting in an
overall mean precision of 69.13 percent.
Unlike previous works, where machines and humans would
address distinct tasks, the current project is unique in that
the investigators recognize that sometimes the jobs requiring
human and/or machine interventions are exactly the same.
Traditionally, in a criminal investigation, a workflow might use
computers to sort pictures and information, then use the people
to identify pictures of interest. In the current approach,
people and machines could sort the images, simultaneously and
people and machines could identify the images. The open problem
is who does what and in which order.
The elasticity of the system is the key feature of the model.
It allows it to adapt to changes in the workload of a task. The
new model allows the machine and human computing resources to
change, when and where needed and without disrupting the
operations.
“Elastically, we would like to decide who is best for a
specific task, or what concentration of people or machines could
be mixed for a specific task,” Blake said. “We also think that
humans could do a first pass of a task, then machines do a
second pass on the same task.”
The model combines the human and machine computing elements
and ignores specific outputs that don’t meet certain quality
requirements. It also sends feedback to improve future
decisions. This feedback allows the model to build a knowledge
base to improve accuracy.
The preliminary results of this study were presented at the
10th IEEE International Conference on Collaborative Computing:
Networking, Applications and Worksharing, where Blake gave the
opening keynote address. The findings will be published by IEEE
press in a paper titled: “Combining Human and Machine Computing
Elements for Analysis via Crowdsourcing.”
The publication was led by Julian Jarrett, Ph.D. student, and
Iman Saleh, research scientist working in Blake’s lab, in the
Department of Computer Science at UM. Other co-authors are Rohan
Malcom and Sean Thorpe, from the University of Technology in
Jamaica; and Tyrone Grandison, from Proficiency Labs in Ashland,
Oregon.
In the future, the researchers would like to recreate
specific past events or create a simulated event, with a staged
“bad guy,” where hundreds of people use their cell phones to
take pictures, then use all the pictures and personal accounts,
to test the new model.