Artificial Intelligence in New Zealand: applications and innovation

Artificial Intelligence (AI) is playing an increasingly significant role in various scientific research areas and real-world applications, ranging from AlphaGo design through medical imaging analysis, earthquake prediction to fish species classification, and fruit maturity estimation to online product recommendation. With world-leading researchers and practitioners, Aotearoa New Zealand is playing an important role in the global AI community. There have been significant achievements in AI in recent years. This special issue aims to highlight recent advances in AI research and developments from the New Zealand community in terms of theory and applications of AI.

This special issue includes ten high-quality manuscripts (Chiewchan et al. 2023, Bi et al. 2023, Babu et al. 2023, Lim et al. 2023, Wilson et al. 2023, Cranefield et al. 2023, Bartlett et al. 2023, Rodger et al. 2023, Sagar et al. 2023, Wang et al. 2023). They cover a wide range of AI techniques and various real-world application areas of AI. AI techniques involved range from traditional AI areas like image analysis and computer vision, natural language processing and multi-agent systems to more recent techniques such as evolutionary machine learning, deep learning, few-shot learning, and explainable AI. These papers also explore how AI can be applied to our daily life, including the primary industries of NZ like agriculture and aquaculture, the critical areas like environment, health and medical, and wellbeing, as well as the considerations of te ao Māori, privacy, transparency, law, social impact in AI.

Agriculture has been significantly impacting the world in various ways and is becoming more critical with the increasingly high food demand caused by the fast population growth. Many traditional methods used by farmers are either too costly in human labour or not sufficiently productive. AI provides great opportunities and potentials for boosting the efficiency and productivity of agriculture in a sustainable and safe way (Talaviya et al. 2020). In this special issue, Chiewchan et al. (2023) develop a multi-agent system for water irrigation in the Canterbury Region of New Zealand, which has the largest proportion of irrigated land (70%) in the country. Water resource consent has been introduced to control water usage, but it can be too expensive and lengthy for farmers with relatively small land to apply. Instead, they can join a community irrigation scheme, but it is hard to accurately estimate how much water they need, since it depends on many factors, such as the type of crop, the size of the farm, any imposed water reduction, and the priority of crops to irrigate. This paper explores a multi-agent system with auction-based negotiation for building an intelligent irrigation management system, to maximise water sharing within a community. Each agent represents a farmer to negotiate with other farmers and make decisions during the buying and selling process. Different auction mechanisms are investigated and the results show the multiunit uniform auction strategy performs the best in effectively distributing excess water in the community.

In 2019, the NZ Government launched its Aquaculture Strategy, which aims to build NZ as a world-leader in sustainable and innovative aquaculture, where AI can play an important role. In this special issue, Babu et al. (2023) and Bi et al. (2023) develop novel computer vision based AI techniques for improving the productivity and reducing human manual efforts in the NZ aquaculture industry. Babu et al. (2023) utilise two convolutional neural network (CNN) methods, Single Shot Detection (SSD) and Faster Regions with CNN (Faster R-CNN), for counting juvenile fish at the NZ Plant and Food Research Limited. Fish counting is a key step in aquaculture breeding and production programmes like in moving fish, evaluating fish density, and estimating required feed amount, but it often requires a large number of human manual hours. In the experiments, by tuning the parameters/settings, SSD and Faster R-CNN can achieve a mean absolute percent errors (MAPE) of 0.1, and MPE-tuned SSD achieved an MAPE of less than 0.05. These results are only slightly higher than the manual count approach, but the computational time is reduced from using up to three hours to approximately 30 seconds to batch processing images to count 100 bins of fish. Bi et al. (2023) propose a new AI-based approach to automatically detect buoys from images taken from a Greenshell mussel farm in the South Island of NZ. Greenshell mussel (Perna canaliculus) is one of the most iconic and economically important aquaculture species in NZ. They grow in farms of more than 20 hectares, where a number of large plastic buoys are connected by ropes, each line extending to several kilometres, which form the backbone structure of the mussel farms. The integrity of these buoys is critical to the mussel growth at sufficient depth and quantities. The proposed AI approach includes data collection, data pre-processing, image segmentation, keypoint detection from images, feature extraction, and classification, where a pre-trained U-Net is used for image segmentation and a new Genetic Programming method is proposed to automatically extract high-level features from keypoints of the images for detecting buoys in the mussel farms. Promising results have been achieved, which further motivate the use of AI in aquaculture.

AI technologies have also been used in addressing various environmental problems, where examples include AI for climate change, biodiversity, conservation, weather forecast, and disaster resiliency (Vinuesa et al. 2020). One of the critical steps in achieving successful global environmental AI is to establish benchmark datasets that can be shared and used by the community as research resources to develop practical AI techniques and tools. In this special issue, Lim et al. (2023) present three image datasets that are collected as part of a government funded project named TAIAO, i.e. Time-Evolving Data Science/Artificial Intelligence for Advanced Open Environmental Science. Many NZ-specific environmental datasets have been collected, and this paper details three image datasets, i.e. images of small invasive mammals, aerial photography of the Waikato region in NZ, and paired aerial photography and satellite imagery that is spatially and temporally consistent. With the goal of building an open-source framework and an openly available dataset repository, the TAIAO project provides a great chance for boosting the development, implementation and evaluation/verification of novel AI techniques for environment.

Health, medicine and wellbeing are another key application area, where AI has been playing an increasingly important role and has a high impact on our work and daily activities (Walsh et al. 2019, Rajpurkar et al. 2022). In this special issue, Wilson et al. (2023) introduce Te Pokapū Hātepe o Aotearoa, the New Zealand Algorithm Hub, which is built to provide a central and secure infrastructure to help NZ's response to the COVID-19 pandemic by hosting and evaluating COVID-19 related algorithms/models. They also summarise the experiences and lessons learned when building the algorithm governance framework for the Hub, in particularly the considerations of ethical concerns, perspectives of Māori as Tāngata Whenua (people of the land), the legal and privacy issues, as well as the algorithms' operational perspectives. The developed governance and evaluation framework could be replicable in other situations involving algorithm management, governance and implementation. This special issue also includes an investigation (Cranefield et al. 2023) on the possible opportunities and barriers when introducing AI-based Digital Productivity Assistants (DPA) as partners to improve people's productivity and wellbeing. A case study is performed on the Microsoft MyAnalytics (MMA) DPA, where machine learning is used to extract work patterns from data of individual worker's activities. Different technology affordances as opportunities, e.g. monitoring work patterns and discovering new goals, and various barriers, e.g. mismatch between categorisation by workers and by DPA, have been identified to co-regulate human work for improving productivity and wellbeing. The findings could potentially be used by other organisations, where workers are with high job demands and high job variety, if using MMA to enhance these workers' the productivity and wellbeing.

With the rapid advances in AI technologies and their real-world applications, a great deal of attention has been recently paid in the non-technical side of AI, such as the ethics, privacy, transparency, security, and social impact of AI (Subramanian 2017, Stahl and Wright 2018). In this special issue, Bartlett et al. (2023) performed a study aiming to understand how exactly the recommendation algorithms work in digital platforms and their implications, since AI based recommendation has been widely used and significantly impacted the behaviours of users, but very little of their technical details are available. A sequence of analyses are performed on the different versions of Privacy Policies and Terms of Use in Spotify (one of the largest digital music service providers) and Tinder (the most popular online dating application), which are very influential but with limited academic scrutiny on their recommendation algorithms. The analysis produces substantive insights about the recommendation algorithms that should be detailed but are currently missed in the Spotify and Tinder's Terms of Use. Another work from Rodger et al. (2023) investigate the use of explainable AI models in assault sentence prediction in NZ. The data is collected from the NZ Legal Information Institute database and converted from PDF to text data. They conduct pre-processing and feature engineering to generate relatively clean data, and then use linear regression to learn an explainable AI model for sentence prediction. Although it is only a proof-of-concept, this study shows the potential of practically using AI in the sentencing process. Finally, this paper also discusses the potential benefits and risks of using AI algorithms in NZ's courts.

An important research topic in AI is to build intelligent systems that are able to cooperatively interact with human in a realistic and real-time manner. Sagar et al. (2023) argue that modelling how cooperation develops in infants can help the development of an AI system with human cooperative abilities. An empirical study is performed to test how infants and their caregiver engage in action-based turn-taking interactions. The results show that infants respond quickly to the actions and interruptions from their caregivers but barely interrupt, which is consistent with the BabyX model, which is an interactive virtual infant created using AI.

Image classification is one of the most popular areas in AI. In this special issue, Wang et al. (2023) conducted a series of experiments to evaluate and compare the use of different feature extractors and classification algorithms in cross-domain few-shot learning for image classification. Different ResNet models pre-trained on ImageNet as feature extractors and several popular classifiers are experimented, and the results show that two classifiers can consistently achieve best accuracy and different feature extractors may be needed at different cross-domain few-shot learning scenarios. Feature vector normalisation and multi-instance learning are also shown to be helpful in improving the accuracy of cross-domain few-shot learning for image classification.

Finally, we would like to express our sincere gratitude to all the reviewers for their expertise and meticulous efforts in ensuring the quality of the papers in this special issue. Last but not least, we also would like to thank Mr Fei He, the Programme Manager–Publishing at Royal Society of New Zealand, for his great support and assistance during the editing process of this special issue.