29 October 2021
About the research project
While there have been great AI advances in image understanding, the most successful approaches require a very large training dataset of annotated images. Obtaining suitable images is often not very difficult, but the bottleneck is in obtaining ground truth annotations, which typically requires human experts. This has been a major inhibiting factor in the application of these techniques in real-world application.
Techniques such as semi-supervised learning and active learning have been suggested to reduce the requirement for labelled data. In both cases the model is initially trained with a small amount of labelled data. The model is then refined by self-labelling additional data or by selecting additional unlabelled images for expert labelling.
Conversely, generative adversarial networks (GANs) may be used to generate additional images for training. In this case, the network is trained to generate labelled synthetic images using only a small dataset for training. This can be a powerful form of dataset augmentation.
This project will investigate the combination of these two complementary approaches. That is, the system will utilise labelled and unlabelled data to build a GAN that may produce additional labelled data. This new data will then be used to refine the model using semi-supervised and/or active learning approaches.
By combining these, and potentially other, approaches, it is hoped that an efficient image understanding workflow can be developed that enables greater application of this transformative technology. One such application is the identification of animal species in camera trap images. Accurate automated identification would allow greater use of this low-cost technique for environmental monitoring. Once developed, the system will be validated by making use of a new camera trap image dataset.
Applicants will be considered for a Research Training Program (RTP) scholarship or Tasmania Graduate Research Scholarship (TGRS) which, if successful, provides:
- a living allowance stipend of $28,597 per annum (2021 rate, indexed annually) for 3.5 years
- a relocation allowance of up to $2,000
- a tuition fees offset covering the cost of tuition fees for up to four years (domestic applicants only)
If successful, international applicants will receive a University of Tasmania Fees Offset for up to four years.
As part of the application process you may indicate if you do not wish to be considered for scholarship funding.
The project is open to domestic (Australian and New Zealand) and international applicants who are already in Australia (onshore) at the time of submitting their application.
Due to current Australian COVID-19 travel restrictions the University cannot accept applications from international applicants who are currently overseas.
Applicants should review the Higher Degree by Research minimum entry requirements.
The project is competitively assessed and awarded. Selection is based on academic merit and suitability to the project as determined by the College.
Additional essential selection criteria specific to this project:
- Demonstrated strong research and analytical skills
- Familiarity with deep learning techniques and tools
Additional desirable selection criteria specific to this project:
- Specific skills in the areas of deep learning with convolutional networks, unsupervised and semi-supervised learning, adversarial networks
There is a three-step application process:
- Select the project, and check you meet the eligibility and selection criteria;
- Contact the Primary Supervisor, Dr Robert Ollington, if you have any questions about the project; and
- Click here to submit an application by the closing date listed above.
- Copy and paste the title of the project from this advertisement into your application. If you don’t correctly do this your application may be rejected.
- As part of your application you will be required to submit a covering letter, a CV including 2 x referees and your project research proposal.
Following the application closing date applications will be assessed within the College. Applicants should expect to receive notification of the outcome by email.