Machine Learning

“Turn Machine Learning into an Engineering Discipline”

Context: Making sense of the ever increasing amounts of data gathered in all areas of human endeavor is an important contemporary challenge.  Machine Learning (ML) technologies address that challenge. Traditionally “machine learning” might refer to symbolic computational approaches arising from the AI community, or statistical pattern recognition. However here it is used more broadly to include allied data-intensive areas such as information retrieval, and computational natural language processing. Machine learning researchers (ultimately and perhaps indirectly) construct software artifacts, and as such we need to be cognizant of aspects of software engineering and protocol design, as well as HCI design.

Problem: Currently ML is largely a technique-driven “black art”. It is hard to compare different methods. Solutions are forced onto problems, rather than being tailored to fit. It is hard to reuse methods. It is often impossible for end users to exploit state of the art methods without the aid of an ML expert. Researchers within ML are often torn by an unresolved tension between further developing the underlying technology (which generates respect of peers) versus helping others solve the real problems by using ML technology. It is very hard to see ML as a field (as a whole) rather than a large bag of tricks.

Opportunity: NICTA proposes to grasp the above tension and turn it into an opportunity and focus for the ML group. ML is an enabling technology – it can contribute significantly to solving challenging real problems, but it is almost never the solution solely in itself. If ML can be made more readily and widely usable, it will have great impact. Making it usable is not merely a matter of packaging existing solutions. ML needs to be viewed from a use-inspired perspective from the beginning. That means there needs to be a focus on problems (clearly articulating what the problem is being solved) cataloging techniques against problems, developing componentized software frameworks that allow the ready deployment of ML technologies and their integration with all of the other information and communication technologies needed to solve large scale and challenging problems. This integration is a common feature of mature engineering disciplines.

Our research plan is to take the above opportunity, view machine learning from a use-inspired perspective, and produce world class research outcomes that make a difference.

Contact:
Leader, Machine Learning Research Group
Bob Williamson
Bob.Williamson<at>nicta.com.au

 Selected Recent ML Publications