School of
Information Technology and Electrical Engineering

Speaker: Ashleigh Richardson
Seminar Date: Thu, 06/02/2020 - 15:00
Venue: 78-311 (Co-Innovation Studio)
Host: Prof Janet Wiles

Seminar Type:  PhD Confirmation Seminar

Abstract: 

Sequence transduction involves converting variable length input sequences into output sequences of another form. While this category of problems is broad, spanning domains such as natural language processing (NLP), security, music, finance, and bioinformatics, almost all tasks that fit within this paradigm are approached using the same machine learning techniques. Neural machine learning techniques for sequence transduction have rapidly advanced in recent years, making full or partial automation of many complex tasks a reality. Neural machine translation (NMT) now allows many people who do not speak a mutual language to communicate effectively. Text-to-speech and speech-to-text systems have altered the ways in which we interact with technology. Music generation and transcription technology now assists many artists in the creative process. Sequence transduction advances have even aided automation of tasks such as protein secondary structure prediction. However, the amount of training data required by many of these techniques often prohibits their use in low-resource settings. Improving low-resource sequence transduction, and making neural techniques more accessible to low-resource tasks, can help to reduce the systemic inequality exacerbated by differences in access to technology.  Transfer learning techniques are emerging to reduce or shift data requirements by inferring additional knowledge from related or unrelated data, but there is a lot of conjecture involved in discriminating between, and applying them. Additionally, it is not clear where the limits of this technology lie, and whether there are ways to further shift or even eliminate some of the data requirements for neural sequence transduction. This work aims to guide research in low-resource neural sequence transduction via four studies. The first three studies investigate the effects of dataset sizes, neural architectures, and relatedness of auxiliary tasks in three transfer learning paradigms for NLP, with a focus on NMT; pre-training, multitask learning, and semi-supervised / unsupervised learning respectively. The final study will generalise these results to other sequence transduction tasks, using information gleaned from earlier studies to guide methodology.

Biography: 

Ashleigh Richardson is a PhD student in the co-innovation group in the School of Information Technology and Electrical Engineering at the University of Queensland. Her supervisors are Prof. Janet Wiles and Assoc. Prof. Marcus Gallagher. Ashleigh received her BE (Hons) in Software Engineering from the University of Queensland in 2018. Her research interests include sequence transduction, natural language processing, machine learning for creativity, and machine learning for low-resource applications. Her research work focuses on guiding the design of low-resource machine learning systems for use in real-world applications.