Environments to Accelerate Machine Learning Based Discovery

  • Goscinski, Wojtek (Primary Chief Investigator (PCI))
  • Abramson, David (Chief Investigator (CI))
  • Padmanabhan, Komathy (Chief Investigator (CI))
  • Ge, Zongyuan (Chief Investigator (CI))
  • Gray, Mark (Chief Investigator (CI))
  • Backhaus, Ann (Chief Investigator (CI))
  • Williams, Allan (Chief Investigator (CI))
  • Carroll, Jake (Chief Investigator (CI))
  • Ward, Nigel (Chief Investigator (CI))
  • Quenette, Steve (Chief Investigator (CI))
  • Bonnington, Paul (Chief Investigator (CI))

Project: Research

Project Details

Project Description

The confluence of big data, Machine Learning (ML) techniques and parallel computing is making AI useful across a range of research areas.. There is increasing sophistication, insight and accuracy which is driving a strong and growing appetite across research groups for access to ML capacity, services, libraries, expertise and training. This project will bring together ML tools, libraries, and access to data, across large GPU deployments nationally. The environment will support core ML tools for preprocessing, annotating, training, and validation, and integrate with software development environments to provide a consolidated platform for ML-based research. A national outreach and training program will engage the researcher community and increase knowledge.

Researchers applying machine learning actively and broadly is a new phenomenon. Therefore, we have collected significant information to ensure that this proposal is evidence based. The proposal is based on the outcomes from an international survey of research groups (referred to as the ML Requirements Survey) across Monash University, University of Queensland and University of Auckland undertaken under “ARDC Discovery Activity: Machine learning infrastructure deployed at scale: understanding requirements, demand, impact and international best practice”. The recommendations and report are available [1].

This project will focus on the key objectives below, which address the most pressing medium-term priorities identified in our researcher survey:
- Create an integrated ML development environment to allow researchers to access interactive development environments (IDEs), data manipulation tools, international ML reference data, ML tools & libraries, SDKs and access to research data, in an efficient and integrated manner, without having to switch between desktop PC, HPC system and data storage;
- Improve system administrator knowledge across HPC centres to build workflows and deploy ML tools and libraries efficiently to improve ML capacity across scarce resources;
- Provide targeted training around the most commonly used data preprocessing and ML tools & techniques (Tensorflow, Keras, PyTorch, SciKit-learn and Caffe are used by 90% of the ML community surveyed); and
- Build Communities of Practice (CoP) around deep neural networks and applied ML domains such as health and HASS, for exchange of knowledge and collaboration, with a particular focus on gathering community requirements across major sites of activity. We will underpin both fundamental ML researchers developing algorithms, and applied ML researchers utilising existing ML techniques to their research.

The project will be driven through strong research engagement and industry collaboration to:
- Accelerate research by providing a friction free environment to test ML ideas, demonstrate success and sometimes ‘fail fast’;
- Promote interdisciplinary research through collaborations with tools and libraries, as well as annotated datasets, and
- Improve the efficiency of usage and best leverage available hardware capabilities.

Application areas underpinned by this project are very broad ranging from neuroscience, clinical science, molecular imaging, and robotics to economics, design and architecture, and HASS.

The platform will be developed across MASSIVE and University of Queensland, and then deployed across NCI and Pawsey. Access to this capability will be available to project partners and nationally through the National Computational Merit Allocation Scheme (NCMAS) across all four sites.
Effective start/end date3/02/203/02/22