Student research opportunities

High-quality/Low False Positive Pattern Recognition utilising Human Assisted Machine Learning

Project Code: CECS_842

This project is available at the following levels:
Honours, Summer Scholar, Masters

Keywords:

Machine Learning, Pattern recognition, Human-Computer interaction, Graphical Models, Active Learning, Preference Elicitation, Quality Assurance

Supervisor:

Dr Christfried Webers

Outline:

Automating Machine Learning processes, such as finding certain objects
in a video stream, is a task that varies in difficulty depending on the
accuracy of the result needed. For example, it can be very easy to
design an algorithm that finds, say, 80% of all occurrences of a
traffic sign in a video stream while producing very few matches where
a traffic sign is recognised because some pattern looks like it but is
in fact not a traffic sign (false positives).
But it becomes increasingly difficult to design algorithms that have
higher detection rates while keeping the number of false positives to
a minimum.

Some real world problems that need to be automated require *very*
high detections rates, so high that the false positive rate of a fully
automated system can not be kept at an acceptable level.
In such circumstances it is sometimes called for to make a manual inspection
of the detection candidates and "weed out" the false positives still present.
In the example above, a highly precise automated system would produce a
map of traffic signs from a video stream, and a number of human operators
would remove the false positives.

This Summer Scholar project aims at coming up with a Quality Assurance (QA) scheme that
minimises the manual effort needed to, confidently, reach a very high
quality. If there is a team of Quality Assurance operators available,
how should the data be passed between them to achieve the highest
quality at the smallest effort? Maybe the skills of the operators are
heterogeneous? Some are fast with low quality, whereas others may be
slow but with high quality output. Therefore, the system also needs to
learn the abilities of the human operators and adapt the data processing network structure accordingly.

Goals of this project

This project will explore a Machine Learning approach to
human assisted computerised quality assurance and develop a
system which adjusts the data flow to the properties learned from
observing the information processing.
The goal is to run experiments with real world data in an advanced
application project currently executed in NICTA.
Time permitted, the implemented scheme may also be used on a
larger "live" problem with a fleet of QA operators.

Requirements/Prerequisites

Strong programming skills (C or C++) are required, Matlab desirable. A background in computer vision and machine learning is desirable.

Student Gain

Experience in research in machine learning and applications to real data.

Contact: