Nordic AI Meet 2021
Norwegian Artificial Intelligence Research Consortium, nora.ai
The Nordic AI young researcher symposium (Nordic AI Meet) represents an important step forward in helping and providing necessary knowledge exchange about AI and its applications in the Nordics. This is particularly important and useful for meeting important challenges in the public sector, industry or civil society regarding the use of AI.
The virtual poster session represents work from young researchers all over the Nordics covering variety of topics linked with AI/ML.
More info: https://nordicaimeet.com
Filter displayed posters (107 keywords)
User Engagement in Gamified Human-Computer Interaction
Bahram Salamatravandi
Physics-informed Neural Network for Viscoelastic Flows
Birane Kane
ComparingBinaryCross-entropyand F1-loss in Multi-labelECG classification
Bjørn-Jostein Singstad, Eraraya Morenzo Muten, Pål Haugar Brekke
It has already been proven that AI can outperform physicians and cardiologists on certain diagnoses based on the ECG, like detecting silent atrial fibrillation. But to be able to replace today’s clinically used, and rule-based ECG interpretation algorithms, the AI-based ECG interpretation algorithm should be able to classify many diagnoses, where multiple diagnoses could be true at the same time. In this study, we train and compare two Convolutional Neural Networks(CNN), one using Binary Cross-entropy (BCE) and one using soft F1-loss.
To train the model we used a dataset containing 88 253 open access 12-lead ECGs with 30 different diagnoses used as ground truth for our supervised model. We trained and validated the CNN models using 10-fold cross-validation and scored the model on the validation split using accuracy, F1-score, Area under the receiver operating characteristic curve (AUROC), and a particular metric developed specifically for the Physionet challenge.
The model, trained using BCE, got an AUROC score of 0.92 ± 0.005 and an accuracy score of 0.97 ± 0.001, while the model trained using F1-loss got an AUROC score of 0.84 ± 0.016 and an accuracy score of 0.95 ± 0.01. However, the BCE model got an F1-score = 0.35± 0.02 and a PhysioNet Challenge score = 0.40 ± 0.02 while the model trained using F1-loss got an F1-score of 0.43 ± 0.02 and a PhysioNet Challenge score of 0.54 ± 0.01.
This implies that the model using BCE loss has the best abilities to predict true positives and true negatives, while the model using F1-loss is better when it comes to avoiding false negatives and false positives. Furthermore, this may indicate that the model using BCE loss is fitted to the given distribution of diagnoses in the used dataset, while the model using F1-loss may generalize better to new data and unknown distributions.
GENDER BIAS IN AI: Perspectives of AI Practitioners
Cathrine Bui & Lara Okafor
Aim: This project explores what perspectives practitioners in AI in Norway have on gender bias in AI by investigating their understanding of technology; how gender bias enters AI systems; and what practices they have in place to detect and address gender bias in AI. Method: Qualitative multiple case studies were conducted. This study interviewed 13 practitioners in the AI field in Norway. Thematic analysis was used to analyze the interviews.
Findings: Practitioners have implemented few practices, most do not use any ethics guidelines, and they delegate responsibilities to other entities. The informants could only identify a few of the entry points of gender bias mentioned by literature, such as biased data, human bias, and a lack of diverse perspectives. The informants with at least one marginalized identity had more knowledge and practices to address gender bias in AI. They were able to identify more systemic causes and higher-impact levers of intervention.
Conclusion: AI practitioners have inherited assumptions and beliefs from predecessors in the AI field on how distancing oneself from one's work achieves neutral objectivity. These beliefs have a significant influence on practitioners' understanding of technology, and as a result, few ethics practices are in place. These assumptions conflate their grasp of what causes gender bias in AI into a technical problem because they underestimate the effects of power. The practitioners see biased data as the main cause, but data is never neutral because no dataset is equally fair for everyone. The practitioners' belief that there exists a form of fairness that will always be correct for everyone at all times without considering the context enables biases to enter AI systems. The AI field needs to examine what technical heritage and taken-for-granted beliefs negatively impact research and practices on gender bias in AI. This study recommends a paradigm shift in practitioners from imagined objectivity to a critical, intersectional perspective that empowers, includes, and creates justice for disadvantaged groups. Inclusion of marginalized perspectives is crucial, and hiring practices should change to increase diversity by training disadvantaged groups in AI.
An Algorithm for Stochastic and Adversarial Bandits with Switching Costs
Chloé Rouyer, Yevgeny Seldin, Nicolò Cesa-Bianchi
Towards an Inclusive Framework for AI-based Care Robots
*Saplacan, Diana; **Martinez, Santiago; *Tørresen, Jim.
References:
[1] J. Fjeld, N. Achten, H. Hilligoss, A. Nagy, and M. Srikumar, “Principled Artificial Intelligence: Mapping Consensus in Ethical and Rights-Based Approaches to Princi- ples for AI,” Social Science Research Network, Rochester, NY, SSRN Scholarly Paper ID 3518482, Jan. 2020. doi: 10.2139/ssrn.3518482.
[2] WHO, “The Right to Health,” Office of the United Nations High Commissioner for Human Rights, Fact Sheet No. 31. [Online]. Available: https://www.ohchr.org/ documents/publications/factsheet31.pdf
[3] D. Saplacan, W. Khaksar, and J. Torresen, “On Ethical Challenges Raised by Care Robots: A Review,” in Proceedings of The IEEE, 20th International Conference in Advanced Robotics and Its Social Impacts (ARSO), Japan/Virtual, 2021, p. 8. DOI: 10.1109/ARSO51874.2021.9542844
[4] V. Dignum et al., “Ethics by Design: Necessity or Curse?,” in Proceedings of the 2018 AAAI/ACM Conference on AI, Ethics, and Society, New York, NY, USA, Dec. 2018, pp. 60–66. doi: 10.1145/3278721.3278745.
[5] N. A. Smuha, “Beyond a Human Rights-based approach to AI Governance: Promise, Pitfalls, Plea,” Social Science Research Network, Rochester, NY, SSRN Scholarly Paper ID 3543112, Feb. 2020. doi: 10.2139/ssrn.3543112.
Not All Comments are Equal: Insights into Comment Moderation from a Topic-Aware Model
Elaine Zosa, Ravi Shekhar, Mladen Karan, Matthew Purver
Domain-Transformation of MRI-derived Time Series with Deep Learning
Erin B. Bjørkeli, John Terje Geitung, Morteza Esmaeili
Recognition of Human Activities using UWB Radar and Deep Learning
Farzan M. Noori
5G NR-based Environment Detection for Seamless Localization utilizing CNN
Ghazaleh Kia, Jukka Talvitie, Laura Ruotsalainen
Unsupervised Learning of Fish as a Novel Class in Detectron2
I-Hao Chen, Nabil Belbachir
Keywords: Unsupervised Learning, Detectron2, Mask R-CNN, Object Segmentation. In this work, we introduce a simple approach to add a novel class, e.g., fish (Atlantic salmon) using the Mask R-CNN [1] algorithm without the need of human annotations. We thereby solve the problem of labelling underwater fish manually, which may be cost- and time-intensive. The main application is to use our algorithm on automated drones like ARV-i [2] in semi-static environments like aquaculture to track fish, estimate their weight or count fish, but other cases are easily adapted since human parameter tuning is minimal. We use the Detectron2 [3] of the Facebook AI Research as implementation of the Mask R-CNN. Our method (see Figure 1) is a sequence of algorithms: Firstly, we split the training data into copies of different sizes. Then the Detectron2 outputs various instance segmentations as inference with low certainty on the data. The different dimensions of the images increase the class variance and detection range – therefore providing more wanted segmentations. Images without segmentations are saved as background files. We curate the data by resizing the segmentations to the original dimensionality of the image. The algorithm now unifies all the potentially overlapping segmentations belonging to one object if they score a high enough IoU (Intersection over Union) value. Afterwards, we apply a series of filters in cascade that check the segmentations for extent, solidity, equivalent diameter, mean value and aspect ratio. That purifies the dataset of misdetections or other foreign objects. These hyperparameters must be tuned by a human as pre-processing step. Finally, the remaining segmentations are pasted randomly on background images. After training on these images, we can input new (or the initial training) data with the novel object and the Detectron2 detects them confidently as novel class (e.g., fish). Current restrictions are that the algorithm stands and fails with the inference on high uncertainty (>95%) and that it needs data where at least some novel objects are separated. We strive to implement a tracking algorithm as well to reduce the detection errors due to occlusions by exploiting a spatiotemporal analysis.
Acknowledgements Acknowledgement to the company SubC3D AS for providing the original video/images taken with an ARV-i [2]. References 1. He K, Gkioxari G, Dollar P et al. (2020) Mask R-CNN. IEEE Trans Pattern Anal Mach Intell 42:386–397. 2. Transmark Subsea (2021) Introducing ARV-i. https://www.transmarksubsea.com/introducing-arv-i/. Accessed 17 Aug 2021 3. Wu Y, Kirillov A, Massa F et al. (2019) Detectron2. https://github.com/facebookresearch/detectron2. Accessed 17 Aug 2021
Automated Noise Monitoring with Machine Learning
Jon Nordby, Erik Sjølund, Ole Johan Aspestrand Bjerke, Fabian Nemazi
The use of Machine Learning is starting to make it possible to automatically detect and classify audio, and there is a growing body of existing research in relevant topics such as Sound Event Detection, Environmental Sound Classification, Acoustic Scene Classification and Universal Source Separation. However, we find that most existing work does not consider the relationship between the noise magnitude/severity (sound level) with noise sources (classification).
Our ongoing research aims to find practical solutions for automatically determining the source of measured noise, by combining sound detection and classification techniques with with sound level measurements according to current noise regulations and acoustical engineering practice.
Our contributions so far have shown that it is feasible to do on-sensor classification of environmental sound on low-cost microcontrollers, that Sound Event Detection (SED) at a known source can be used to create logbooks of noisy activity, and that doing SED at both source and receiver allows to determine the source of noise experienced at receiver.
Current work focuses on new task formulations for Machine Learning that incorporate the requirements of acoustic modelling and noise regulations, including Noise Detection \& Classification. We hope that the formalization of these tasks will enable more research and increased relevance for understanding acoustical noise using Machine Learning.
Our primary method of research involves case-studies and demonstrator projects in the real-world, and we invite parties interested in noise, acoustics and machine learning to collaborate with us.
Design of trustworthy and inclusive AI services in the public sector
Karolina Drobotowicz
My proposed research focuses on how trustworthy and inclusive AI can be effectively implemented in the public sector: understanding the technological needs, challenges and regulations for devising trustworthy AI services, while developing tools, methods and critical assessment of outcomes. It builds on research in AI ethics including transparency, inclusion, accountability, auditability, and explainable AI, while engaging aspects of HCI and Human-AI interaction.
I am using a multidisciplinary and participatory approach in this research. On one side, I am collaborating with key stakeholders, such as public sector representatives (eg. from the City of Helsinki), public services designers and developers, policy and legal experts. I plan to conduct qualitative interviews, focus group sessions, and if possible, ethnographic and case studies with this group. The work has already started as the study on AI Act implications for educational and public services. On another side, I am including civil society in the study, such as regular citizens and representatives of vulnerable communities that might be affected by the public AI services. With this group, I have been conducting interviews, workshops and focus groups. Finally, I intend to gather both groups to perform the co-design sessions. As the result, I intend to publish: 1) frameworks of the successful participatory method for public AI services and 2) exemplar trustworthy and inclusive public AI service interface.
In summary, I aim to provide realistic solutions and recommendations for making public AI services trustworthy and beneficial to society, thereby under-represented groups. This could contribute, first, to the empowerment of citizens by acknowledging their experiential expertise and providing them ways of participation in developing public AI services. Second, to the successful implementation of AI services by the public services providers. Third, to understanding the implications of and preparations for the AI Act.
References: [1] European Commission, “Proposal for a Regulation laying down harmonised rules on artificial intelligence, ”https://digital-strategy.ec.europa.eu/en/library/proposal-regulation-laying-down-harmonised-rules-artificial-intelligence, accessed 28.05.2021 [2] Artificial Intelligence Governance and Auditing (AIGA) project, https://ai-governance.eu/, accessed 28.05.2021 [3] Young, M., Magassa, L. & Friedman, B. Toward inclusive tech policy design: a method for underrepresented voices to strengthen tech policy documents. Ethics Inf Technol 21, 89–103 (2019) [4] Drobotowicz K., Kauppinen M., Kujala S. (2021). Trustworthy AI Services in the Public Sector: What Are Citizens Saying About It?. In: Requirements Engineering: Foundation for Software Quality, REFSQ’21
Predicting progression & cognitive decline in amyloid-positivepatients with Alzheimer’s disease
Hákon Valur Dansson, Lena Stempfle, Hildur Egilsdóttir, Alexander Schliep, Erik Portelius, Kaj Blennow, Henrik Zetterberg, Fredrik D. Johansson
A cohort of n=2293 participants, of whom n=749 were Aβpositive, was selected from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database to study heterogeneity in disease progression for individuals with Aβpathology. The analysis used baseline clinical variables including demographics, genetic markers, and neuropsychological data to predict how the cognitive ability and AD diagnosis of subjects progressed using statistical models and machine learning. Due to the relatively low prevalence of Aβpathology, models that fit only to Aβ-positive subjects were compared to models that fit an extended cohort including subjects without established Aβpathology, adjusting for covariate differences between the cohorts.
Aβpathology status was determined based on the Aβ42/Aβ40ratio. The best predictive model of change in cognitive test scores for Aβ-positive subjects at the two-year follow-up achievedanR2score of 0.388 while the best model predicting adverse changes in diagnosis achieved a weightedF1score of 0.791. Conforming to expectations, Aβ-positive subjects declined faster on average than those without Aβpathology, but the specific level of CSF Aβwas not predictive of progression rate. When predicting cognitive score change four years after baseline, the best model achieved anR2score of 0.325 and it was found that fitting models to the extended cohort improved performance. Moreover, using all clinical variables outperformed the best model based only on a suite of cognitive test scores which achieved anR2score of 0.228.
Our analysis shows that CSF levels of Aβare not strong predictors of the rate of cognitive decline in Aβ-positive subjects when adjusting for other variables. Baseline assessments of cognitive function account for the majority of variance explained in the prediction of two-year decline but are insufficient for achieving optimal results in longer-term predictions. Predicting changes both in cognitive test scores and in diagnosis provides multiple perspectives of the progression of potential AD subjects.
Detection of events in real-time news streams with dark entities using knowledge graphs
Marc Gallofré Ocaña, Andreas L. Opdahl
Monitoring melt by classifying Cryosat-2 waveforms
Martijn Vermeer, David Völgyes, Malcolm McMillan, Daniele Fantin
Allocating Opportunities in a Dynamic Biased Mode
Meirav Segal
We present a model of a dynamic allocation process to simulate a scenario of college admission. Higher education is a key element towards many career paths. As such, access to higher education could be crucial for self fulfillment and/or financial security. Unfortunately, currently there are groups with lessened access to this opportunity due to societal biases.
We examine the dynamics of a decision process with two groups in the population, defined based on a protected attribute (such as race or gender). While the innate ability is uniformly distributed in both groups, one group is disadvantaged in the sense that there is societal bias against its members, which leads to reduced success probability. The admission decisions affect not only the DM's utility (sum of discounted rewards), but also future bias, which in turn affects the range of possible future utility. Since there is a relation between bias and utility, it is interesting to observe how the utility maximizing policy behaves under different model parameters.
We consider two possible feedback mechanisms. Admission policy can increase group qualification by setting a lower acceptance bar for the disadvantaged group (i.e. affirmative action). This act could potentially generate more role models and provide investment incentive, since there is a greater chance of being admitted. On the other hand, this very action might lead to a reversed effect, as lowering the bar means that less qualified members of this group are admitted. Hence, less admitted individuals from that group are likely to graduate. This might increase societal biases by reinforcing existing stereotypes and internalizing them by members of that group.
Our preliminary analysis for one type of bias shows that there are three regions of the bias state with distinct behaviour of the utility maximizing policy - granting none of the opportunities to the disadvantaged group, setting the same bar for both groups and applying affirmative action. These regions are also clearly seen in experimental results produced using policy iteration. In order to compare policies with respect to fairness, we define the notion Horizon Fairness which considers the impact of the policy on future bias. In the future, we will extend our analysis to the other kind of bias and the combination of both. In addition, we would like to provide some results for optimization under uncertainty of the feedback mechanism.
AI-enabled proactive mHealth
Muhammad Sulaiman, Anne Håkansson, and Randi Karlsen
Reinforcement Learning for Optimal a Hour-Ahead Electricity Trading with Battery Storage
Peyman Kor
I know where you are going! Transport mode detection and understanding feature importance based on smartphone sensors
Philippe Büdinger and Tor-Morten Grønli
NLC activity detection using Convolutional Neural Network
Rajendra Sapkota, Puneet Sharma, Ingrid Mann
Complexity and Predictability Analysis of the Elder Problem Using Big Data and Machine Learning
Roman Khotyachuk, Klaus Johannsen
1. Investigated the steady-state solutions of the Elder problem with regards to the Rayleigh numbers (Ra), grid sizes, perturbations, etc. 2. Analyzed the complexity of solutions regarding time, solution types, and other factors. 3. Created a tool for visual exploration of large solution sets from the Elder problem. 4. Developed predictive models for the Elder problem using different classification methods.
Our predictive models can be divided into the following types, depending on how the predictors (features) are designed: 1) fully informed models; 2) partially informed models; 3) “black box” models. The best of our models can predict a steady-state of the Elder problem (i.e., when time t > 50 years) with 95% accuracy at t=8-9 years.
Improved Analysis of the Tsallis-INF Algorithm in Stochastically Constrained Adversarial Bandits and Stochastic Bandits with Adversarial Corruptions
Saeed Masoudian, Yevgeny Seldin
Predicting Grades of Brain Tumor from Histology Images Using Convolutional Neural Networks
Saruar Alam, Alexander S. Lundervold, and Arvid Lundervold
Gliomas are the most common type of primary brain tumor and are characterized by large morphological and genetic heterogeneity [HeteroGen], varying degree of tumor growth, speed of spread and tumor recurrence, poor prognosis, and high lethality. Gliomas are classified into low-grade glioma (LGG;grade I and II) and high-grade glioma (HGG;grade III and IV), and treatment for a glioma depends on its grade. Grading is performed by histological examination of (representative) glioma tissue according to several phenotypic characteristics that describe cell activity and tumor aggressiveness, usually requiring an experienced neuropathologist. The knowledge of tumor grade together with other factors such as age, clinical condition and tumor location, assists in treatment planning, and helps estimating prognosis and expected survivability. For instance, patients with a grade IV glioblastoma will have an average survival time of 12-18 months from the date of diagnosis. Automatic glioma grading, referred to as tumor grading, is desirable as it reduces inter and intra operator variability in determining tumor grade in histological images from a patient. Deep learning models used for image classification, including tumor grading, are often based on convolutional neural networks (CNN). This study employs two different pre-trained CNN models, EfficientNet and ResNet, to classify tumor grades.
Material and methods
We collected histology images with grades II, III, and IV from the TCGA-GBMLGG project, as used in [PathomicFus] and [HistoGeno] (https://hub.docker.com/r/cancerdatascience/scnn). In this setting, a histological image comprises a single region of interest (ROI) from whole slide imaging (WSI) scans of paraffin embedded sections from a glioma patient. Multiple ROIs can be extracted from the WSI slide, giving the possibility that several histology images are linked to a given patient. Our dataset contains 1458 histology images (resolution: 1024x1024) from altogether 736 patients divided into Grade II:181, Grade III:205,and Grade IV:350 subjects. We used an EfficientNet-B0 model for training, selected from the EfficientNet's family of B0-B7, as it has been shown to reduce training and inference time, with fewer learnable parameters, without compromising performance. EfficientNet contains several memory-efficient Inverted Residual Blocks (MBConv). We compared our trained model with Pathomic-fusion~[PathomicFus](https://github.com/mahmoodlab/PathomicFusion) and variants of ResNet models:ResNet-18 and ResNet-34 by using receiver-operating-characteristic-area-under-the-curve (ROC-AUC) and F1-Score, with fifteen cross-validations.
Results
Figure 3 (in the poster) shows tumor grading results from six subjects. The upper row is histology images with ground truth labels. The color-coded images in the lower row depict gradient-weighted class activation maps (Grad-CAM),i.e.spatial location of image information most important for the tumor grade prediction.
Discussion
We adopted pre-trained EfficientNet-B0 for tumor grading classification and compared the performance with Pathomic-fusion and two ResNet variants. The proposed model achieves higher accuracy in its tumor grade predictions (cf. Fig. 2 in the poster). In further work we will expand and evaluate our approach for obtaining visual explanations for the model's predictions and also incorporate estimates of model uncertainty. Incorporation of multiparametric brain MRI from the same subjects is also a challenge to address.
References
[HeteroGen] R. Chow et al. Am J Roentgenology 2018(doi:10.2214/AJR.17.18754) [PathomicFus] R.J. Chen et al. IEEE TMI 2020(doi:10.1109/TMI.2020.3021387) [HistoGeno] M.Pooya. et al. PNAS 2018;115(13):E2970-E2979(doi:10.1073/pnas.1717139115)
Evaluating Artificial Intelligence Explanations on Domain Experts
Steven Hicks, Cise Midoglu, Inga Strumke, Steffen Mæland, Andrea Storås, Malek Hammou, Pål Halvorsen, Vajira Thambawita, and Michael Riegler
Classification of deforestation alerts
Tord Kriznik Sørensen
The suggested methodology is semantic segmentation, starting with a U-Net-like architecture. Multispectral images from the freely accessible Sentinel-2 satellites (S2) will be used. S2 provides 10m, 20m and 60m resolution optical images with around 7 days update frequency. The GLAD alerts, and their immediate surrounding area, will be manually labelled, building up an S2 – GLAD alert – primary driver dataset. This dataset will be created in an iterative fashion in collaboration with domain experts.
In addition to the imagery and the labels, the input data is also enriched with elevation models, which helps to differentiate classes. E.g. rivers always flow in gradient direction, while roads do not.
Finally, only a tiny fraction of the GLAD alerts are labeled, but a huge amount of them are available from previous years. Building on the unlabeled dataset, a semi-supervised training scheme can be implemented.
Many of the deforestation activities are illegal and often local governments don’t have the resources for real time monitoring. International NGOs and environmental monitoring programs can provide alerts to local authorities in a timely manner. Identifying the primary drivers is an important step to determine the seriousness of the alerts and prioritize which to act upon.
The project lasts from October 2021 to June 2022.
Modeling Risky Choices in Unknown Environments
Ville Tanskanen, Chang Rajani, Homayun Afrabandpey, Aini Putkonen, Aurélien Nioche, Arto Klami
Chebyshev-Cantelli PAC-Bayes-Bennett Inequality for theWeighted Majority Vote
Yi-Shan Wu, Andrés R. Masegosa, Stephan S. Lorenzen, Christian Igel, Yevgeny Seldin
A Dynamic and Self-Optimizing Approach for Job-Shop Floor Path Finding and Collision Avoidance in Industry 4.0
Yigit Can Dundar
Modeling Piano Nonlinearities Using Deep Learning
Riccardo Simionato