Deep Learning research projects

Learning Deep Semantics for Automatic Translation between Human Languages (ARC DP, 2016 - 2019)

(ARC DP, 2016-2019)

Project lead: Prof. Reza Haffari

The modern world relies increasingly on automatic translation of human languages to deal with billions of documents. Current translation systems struggle on complex texts often producing misleading or incoherent outputs. Sentences are translated independently ignoring their overall, document-wide, context. We are addressing these issues by developing a new approach using semantics - the underlying meaning of the text - to drive translation; both as discrete structures and continuous representations learned via deep learning. Our work will improve translation quality and lead international research and industry efforts in translation, improving the end-user experience.

Read the full project report

Deep Learning for Cybersecurity

(DST/Data61, 2017-2021)

Project lead: Prof. Dinh Phung

The meteoric rise of deep learning is in part due to an ability to learn feature representations and complex non-linear structure in datasets. Particular successes in vision, speech and natural language, all exhibit hierarchies of patterns at fine to coarse scales. Cyber is ready for similar success owing to its complex, non-linear detection tasks. The goal of this project is to develop new deep learning techniques for cybersecurity with a focus on software vulnerability detection.

Generative Adversarial Networks for Behaviour Discovery

(DST, 2018-2019)

Project lead: Prof. Dinh Phung

The aim of this collaborative project is to assess the viability of a range of artificial techniques for discovering novel behaviour for teams and forces (teams of teams) in the context of adversarial multi-agent simulations.

Towards Robust Learning Systems via Amortised Optimisation and Domain Adaptation

(DST/Data61, 2018-2019)

Project lead: Dr. Trung Le and Prof. Dinh Phung

This project aims to develop deep learning, optimisation and domain adaptation methods towards robust and trustworthy machine learning systems.

Learning to Learn with Less Labels

(Defence Advanced Research Project Agency, 2019-2023)

We provide a unified treatment of learning from less labels which is based on learning to learn (L2L), and addresses fundamental shortcomings of previous approaches to this framework. L2L aims to learn inductive biases from a collection of tasks in order to apply these to learn new tasks more effectively. Current approaches do not allow for fully autonomous agents, which may continually encounter new problems and tasks, while operating in an environment with dynamic change of a nature that is not foreseen by the agents’ designer.