Nordic Artificial Intelligence Research and Development - 4th Symposium of the Norwegian AI Society, NAIS 2022, Oslo, Norway, May 31 - June 1, 2022, Revised Selected Papers

Anis Yazidi, Evi Zouganeli, Gustavo Borges Moreno e Mello, Pedro G. LindNais 2022 · DOI · DBLP

Abstract

The 2022 Symposium of the Norwegian AI Society (NAIS) showcased a vibrant snapshot of the Nordic AI landscape, emphasizing a pragmatic shift toward real-world application. This paper summarizes key contributions in knowledge infusion, weakly supervised learning, and specialized applications in healthcare, industrial maintenance, and cybersecurity. The proceedings highlight a concerted effort to move AI from controlled laboratory settings into the complex environments of hospitals, power grids, and digital defense systems.

Introduction: Addressing the Deployment Gap

A central theme of NAIS 2022 was the "deployment gap"—the discrepancy between high performance on curated benchmarks and the challenges of messy, unpredictable real-world environments. Researchers identified three primary hurdles:

  1. Data Scarcity: High-quality labeled data is often prohibitively expensive or unavailable in specialized fields like embryology or cybersecurity.
  2. Environmental Complexity: Autonomous systems require "cognitive architectures" that integrate safety, ethics, and intuition beyond simple task execution.
  3. Data Heterogeneity: Real-world data is frequently tabular, incomplete, and heterogeneous, resisting standard deep learning approaches.

Methodological Frameworks

The symposium presented a diverse toolkit designed to handle these constraints:

Knowledge Infusion and Weak Supervision

To mitigate the cost of retraining large models, Biermann et al. proposed an ensemble-based architecture that "infuses" expert knowledge into neural networks. By combining raw input embeddings with outputs from specialized "expert" models, downstream tasks benefit from domain-specific insights at a fraction of the computational cost. Similarly, researchers utilized frameworks like Snuba and Snorkel to generate "weak labels" from heuristics, enabling effective fake news detection even when large-scale ground truth is missing.

Specialized Statistical Modeling

For medical phenotyping, Generalized Low-Rank Models (GLRM) were employed to handle diverse data types (binary, ordinal, and count data) within complex medical questionnaires. In industrial settings, the Matrix Profile (MP) algorithm proved effective for time-series anomaly detection, offering a robust, parameter-light method for identifying motifs and outliers in sensor data.

Computer Vision in Specialized Domains

Standard architectures like YOLO v5 and ResNet were adapted for niche medical tasks. Notable applications included tracking human embryo cleavage stages and the unsupervised clustering of video frames from Intracytoplasmic Sperm Injection (ICSI) procedures.

Datasets and Benchmarking

The research leveraged several specialized datasets:

  • Medical/Biological: Norwegian cervical cancer screening data (n=6359) and a novel dataset of 890 time-lapse videos of human embryo development.
  • NLP and Social Media: The Switch-Board dialect corpus (SwDA) and the NELA-GT-2019 dataset for fake news detection.
  • Synthetic Data: A new simulator for Norwegian university admission data was introduced to facilitate "fair matching" algorithm testing while preserving student privacy.

Key Results and Insights

The symposium reported significant performance metrics across various domains:

  • Medical Imaging: The YOLO v5-based embryo tracking system achieved a Matthews Correlation Coefficient (MCC) of 0.77, with peak performance during the critical 2-cell stage.
  • Fake News Detection: A RoBERTa-based model utilizing weak supervision reached 77.9% accuracy, outperforming purely supervised methods on limited datasets.
  • Tabular Data Repair: The SINATRA approach, which uses knowledge graphs to "heal" broken datasets, achieved an F1-score of 1.0 on several benchmark missing-value tasks.
  • Visualization: The introduction of the Kernelized Taylor Diagram (KTD) provided a robust method for visualizing similarities between data populations, accounting for non-linear relationships and outliers more effectively than traditional Taylor Diagrams.

Discussion: The Future of Nordic AI

A recurring insight was the "Right Brain" of robotics: the argument that social integration for robots requires moving beyond logic toward holistic perception and moral evaluation. Furthermore, the "sequential paradigm" in NLP—appending expert knowledge to sequences—was found to enhance explainability by allowing researchers to observe model attention on specific expert advice.

However, challenges remain regarding generalization. Many models were developed using data from single clinics, raising questions about their performance across different hardware or demographics. Future research is trending toward "Multi-Expert" systems and refined "Survival Analysis" to predict the timing of medical events rather than just their occurrence.

Conclusion

NAIS 2022 demonstrated that the Nordic AI community is successfully transitioning from theoretical exploration to practical problem-solving. By focusing on data efficiency, explainability, and specialized domain knowledge, these researchers are paving the way for AI that is not only powerful but also reliable and deployable in critical real-world sectors.

Summary updated

BibTeX
@inproceedings{yazidi2022nordic,
  author = {Anis Yazidi and Evi Zouganeli and Gustavo Borges Moreno e Mello and Pedro G. Lind},
  title = {{Nordic Artificial Intelligence Research and Development - 4th Symposium of the Norwegian AI Society, NAIS 2022, Oslo, Norway, May 31 - June 1, 2022, Revised Selected Papers}},
  booktitle = {Nais},
  year = {2022},
  doi = {10.1007/978-3-031-17030-0},
}

Loading related content...

Figure