Introduction to the special issue on Innovative Applications of Artificial Intelligence (IAAI 2023)

This special issue of AI Magazine covers select applications from the IAAI conference held in 2023 in Washington, DC. The articles address a broad range of very challenging issues and contain great lessons for AI researchers and application developers. The goal of the Innovative Applications of Artificial Intelligence (IAAI) conference is to highlight new, innovative systems and application areas of AI technology and to point out the often-overlooked difficulties involved in deploying complex technology to end users. Those of us who have ventured out of the realm of pure research and tried to build applications to be used by our fellow humans realize that it takes a lot more than just brilliant algorithms to make an application survive in the real world. Each application that succeeds is worth celebrating, and the teams behind them are due wholehearted congratulations. It is in this spirit that we bring you this special issue covering select applications from the IAAI conference held in February 2023 in Washington, DC. The articles address a broad range of challenging issues and contain lessons for fellow AI researchers and application developers. IAAI acceptance criteria are different from most academic conferences in that the end-user application must come first and foremost. A paper written for the annual AAAI or IJCAI conferences is unlikely to be accepted for IAAI because these papers focus on the innovation in the algorithm. IAAI focuses on how to get algorithms to the end user. A paper that describes a small change in a learning model to achieve 1% improvement in accuracy over related work is not appropriate for IAAI. Meanwhile, IAAI would be very interested in a similar paper saying that the current model is not deployable (e.g., due to size or training data), but a small change in the model that loses 1% accuracy allows it to be deployable. The articles in this issue cover humanitarian needs, manufacturing, and forecasting. A common theme is that all deployed applications work directly with end users to design a system that meets end-user needs. Many of the papers have co-authors from the end user community, which strengthens the paper significantly. The papers focus on end-user concerns, both in terms of solving the true end-user problem and in terms of generating explainable results. The first article by Rahul Nair from IBM with Bo Madsen and Alexander Kjærum from the Danish Refugee Council presents a system that forecasts the dynamics of refugee displacements. The system, Foresight, supports long-term forecasts aimed at humanitarian response planning. The explainable system provides evidence and context supporting the forecast and allows analysts to explore “what if” scenarios. Challenges to fielding this system include human-centered design, acceptance in the user community, and technical maturity, notably the lack of high-quality data. Foresight now covers 25 countries and 89% of all displaced populations globally. Shresth Verma and colleagues from Google, Harvard, Purdue, and the Indian Institute of Technology worked with ARMMAN, a nongovernmental organization to support maternal health care. Health care workers reach out to mothers to boost engagement with the health care services; these workers have limited availability compared to the number of beneficiaries. SAHELI helps identify the best recipients for service calls, prevented a drop in engagements by 30.5%, and is on track to serve one million beneficiaries by the end of 2023. This scale and impact have been achieved through multiple innovations in the model and its development, in preparation of real-world data, in deployment practices, and through careful consideration of responsible AI practices. ZhaoYang Zhu and colleagues from Alibaba deployed a system to forecast electricity load and can handle extreme weather conditions such as high temperatures or hurricanes. Accurate forecasting leads to more reliable and safe planning for the power grid. eForecaster contains a suite of explainable algorithms for diversified energy forecasting, leading to a 40% improvement in error, reduced manual work, and increased user acceptance through explainable guidance. The article describes four applications that have been deployed in seven provinces in China. Mihye Kim and her colleagues from Hyundai Capital and the Korean Advanced Institute of Technology (KAIST) in South Korea developed a system to predict the residual value of a vehicle over time. This information is used to determine credit lines and leasing rates to reduce revenue loss, reduce cases of loan default, and prevent fraud. The system also helps buyers avoid disreputable car dealers: these price estimates are shared on a used car website by the South Korean government. The paper describes real-world operational requirements such as compliance with regulations, and generalization to unseen input, for example, new and rare car models. Kyung Pyo Kang and colleagues from Kyung Hee University and Hyundai Motor in South Korea look at novel manufacturing design: products must be innovative, from the perspective of end-user preferences and from the perspective of competition and copyright infringement. A design infringement that is detected late in the manufacturing process might cost over $3 million US dollars. Their system, in use at Hyundai Motor, performs similarity verification of wheel designs and shortened the verification time by 90% to a maximum of 10 min. Designers no longer need to manually search for similar wheel images and can instead focus on other important aspects of the design process and bring new products to market faster. Chao Zheng, Xu Cao, and their colleagues at Tencent and New York University (NYU) looked at the challenge of high-definition (HD) maps for autonomous vehicle navigation. Their Tencent HD Map AI (THMA) system helps analysts process centimeter-resolution laser image datasets and label images. The active learning approach serves over 1000 labeling workers and generates more than 30,000 km of HD map data per day. With over 90% of HD map data labeled automatically by THMA, the system accelerates traditional HD map labeling processes by more than tenfold, significantly reducing manual annotation burdens and paving the way for more efficient HD map production. Rabia Ali and colleagues from Endress+Hauser in Germany and the Norwegian University of Science and Technology developed an approach to detect the weld seam between two metals. Current computer vision approaches are unable to meet the stringent tolerances required for industrial use, leading to significant human intervention to verify and correct every detected edge. Combined with a prefiltering approach that eliminates anomalous workpieces, their system can be deployed directly on laser welding machines, thus saving significant production time and cost. In a similar vein, Hayden Gunraj and his colleagues at the University of Waterloo, DarwinAI, and Moog use computer vision for quality inspection of solder joints. Solder joint defects affect a variety of printed circuit board components. The current manual inspection process is time-consuming and error-prone. SolderNet is an explainable computer vision algorithm that leads to high-throughput, high-performance, and zero-fatigue inspection. SolderNet has been used for over 26 million inspections with an overall escape rate below 0.01%. This selection of articles contains only a few of the interesting articles presented at IAAI 2023. We encourage our readers to look at the proceedings of IAAI 2023 and to submit papers of their own deployed systems to future IAAI conferences. The authors declare that there is no conflict. Karen Z. Haigh is a Consultant and Speaker in the area of Cognitive Radio Frequency systems, and is the Co-author of the book “Cognitive Electronic Warfare: An Artificial Intelligence Approach.” Alexander Wong is a Professor at the University of Waterloo, the Canada Research Chair in Artificial Intelligence and Medical Imaging, and Co-founder of DarwinAI. YuHao Chen is a Research Assistant Professor at the University of Waterloo focusing on vision and image processing.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.