Back to research

AI deployment and systems design

Overview

While excitement about the potential of artificial intelligence (AI) technologies continues to build, a gap is emerging between our aspirations for the benefits of AI and our ability to deploy these technologies to tackle real-world challenges. Achieving the full potential of AI – and its benefits for society and the economy – requires the ability to safely and effectively deploy AI systems at scale. Research in this theme considers the advances in AI and system design that can manage the complex interactions that arise in real-world applications, from innovations in statistical emulation to software engineering for machine learning deployment.

What advances are needed to ensure that AI systems operate safely and effectively in deployment?

A significant proportion of attempted AI deployments fail, highlighting a suite of practical issues that arise when trying to integrate AI into real-world systems – from data management and use, to model performance, to user experience. These failures not only hold back the economic potential of AI, they also expose individuals, communities and societies to new forms of harm. Our AI deployment and system design research theme considers the interventions that can support safe and effective AI systems in real-world contexts.

Research in this theme considers:

systems and software engineering for machine learning deployment, investigating how new software and data architectures can support system adaptability, scalability, and autonomy;
statistical emulation and hierarchical probabilistic modelling, exploring the role of emulation and simulation in improving AI system performance;
what advances in modelling and system design can improve accuracy, robustness, safety and effectiveness, including real-time inference and decision-making; and
AI applications in healthcare, science, and industry.

Projects

AutoAI

Climate Ensembling

General Circulation Models (GCMs) of Earth's climate provide robust simulations of large-scale average climatic variables, such as end-of-century global average temperature, under various future greenhouse gas emissions scenarios. Translating these outputs to insights that can be used to manage the local-level impacts of climate change is challenging, and requires innovations in modelling, data management, and software engineering.

Challenges in machine learning deployment

In addition to being a thriving academic discipline, machine learning is increasingly adopted as a solution to real world business problems. But underneath this seeming success lies a chasm of failures. This project studies reports of ML deployment, and investigates issues practitioners face at each step of the ML deployment pipeline

Data-Oriented Architectures for AI-based Systems

Data Oriented Architecture (DOA) is a software architecture pattern that creates data-driven, loosely coupled, decentralised, and open systems. DOA achieves these goals by exposing systems' data as first class citizen to distributed, stateless, and asynchronous systems’ components. These design decisions enable DOA-based systems to achieve desirable properties such as data availability, reusability, and monitoring, as well as systems adaptability, scalability, and autonomy.

Evaluation of dataflow for ML deployment

When deploying machine learning algorithms in real-world systems, software developers face a new set of challenges. This project asks: "What are the underlying reasons for these challenges to arise?", and proposes that the problem lies with the modern trend of building software systems as microservices.