Data flow analyses are instrumental to effective compiler optimizations, and are typically implemented by extracting implicit data flow information from traversals of a control flow graph intermediate representation. The Regionalized Value State Dependence Graph is an alternative intermediate representation, which represents a program in terms of its data flow dependencies, leaving control flow implicit. Several analyses that enable compiler optimizations reduce to NP-Complete graph problems in general, but admit linear time solutions if the graph’s treewidth is limited. In this paper, we investigate the treewidth of application benchmarks and synthetic programs, in order to identify program features which cause the treewidth of its data flow graph to increase, and assess how they may appear in practical software. We find that increasing numbers of live variables cause unbounded growth in data flow graph treewidth, but this can ordinarily be remedied by modular program design, and monolithic programs that exceed a given bound can be efficiently detected using an approximate treewidth heuristic.

Multi-core and GPU-based systems offer unprecedented computational power. They are, however, challenging to utilize effectively, especially when processing irregular data such as graphs. Graphs are of great interest, as they are now used to model geographic-, social- and neural networks. Several interesting programming frameworks for graph processing have therefore been developed these past few years.
In this work, we highlight the strengths and weaknesses of the Galois, GraphBLAST, Gunrock and Ligra graph frameworks through benchmarking their single source shortest path (SSSP) implementations using the SuiteSparse Matrix Collection. Tests were done on an Nvidia DGX2 system, except for Ligra, which only provides a multi-core framework. D-IrGL, built on Galois, also provided a multi-GPU option for SSSP. We also look at program size, documentation and overall ease of use.
High performance generally comes at the price of high complexity. D-IrGL shows its strength on the very largest graphs, where it achieved the best run-time, while Gunrock processed most other large sets the fastest. However, GraphBLAST, with a relatively low- complexity interface, achieves the greatest median throughput across all our test cases. This despite that its SSSP implementation size is only 1/10th of Gunrock, which for our tests has the highest peak throughput and the fastest run-time in most cases. Ligra had less computational resources available, and consequently performed worse in most cases, but it is also a very compact and easy to use framework. Futher analyses and some suggestions for future work are also included.

This paper investigates the problem of finding large prime gaps (the difference between two consecutive prime numbers, pi+1 – pi) and on the development of a small, efficient program for generating such large prime gaps for a single computer, a laptop or a workstation.

The Shallow Water Equations model the fluid dynamics of deep ocean flow, and are used to simulate tides, tsunamis, and storm surges. Numerical solutions using finite difference methods are computationally expensive enough to mandate the use of large computing clusters, and the cost grows not only with the amount of fluid, but also the duration of the simulated event, and the resolution of the approximation. The benefits of increased resolution are mostly connected to regions where complex fluid interactions occur, and are not required globally for the entire simulation. In this paper, we investigate the potential for conserving computational resources by applying Adaptive Mesh Refinement to dynamically determined areas of the fluid surface. We implement adaptive mesh refinement in a MacCormack finite difference solver, develop a performance model to predict its behavior on large-scale parallel platforms, and validate its predictions experimentally on two computing clusters. We find that the solver itself has highly favorable parallel scalability, and that the addition of refined areas introduces a performance penalty due to load imbalance that is at most proportional to the refinement degree raised to the third power.

The Norwegian National Strategy for Artificial Intelligence (NNSAI) published in 2020 includes seven principles of ethical AI. This paper explores whether those seven principles are stated in a clear enough way and are feasible to be satisfied by a specific AI system. We build an implementation of the Gale-Shapley algorithm to allocate kindergarten places in Bergen, Norway. The presented solution is then evaluated against the ethical principles from the NNSAI. We argue that it is difficult to respect all the ethical principles when implementing a solution to a matching problem.

The use of artificial intelligence for decision making raises concerns about the societal impact of such systems. Traditionally, the product of a human decision-maker are governed by laws and human values. Decision-making is now being guided - or in some cases, replaced by machine learning classification which may reinforce and introduce bias. Algorithmic bias mitigation is explored as an approach to avoid this, however it does come at a cost: efficiency and accuracy. We conduct an empirical analysis of two off-the-shelf bias mitigation techniques from the AIF360 toolkit on a binary classification task. Our preliminary results indicate that bias mitigation is a feasible approach to ensuring group fairness.

Traditional supervised learning requires significant amounts of labeled training data to achieve satisfactory results. As autonomous perception systems collect continuous data, the labeling process becomes expensive and time-consuming. Active learning is a specialized semi- supervised learning strategy that allows a machine learning model to achieve high performance using less training data, thereby minimizing the cost of manual annotation. We explore active learning for autonomous vehicles, and propose a novel deep active learning framework for object detection and instance segmentation. We review prominent active learning approaches, study their performances in the aforementioned computer vision tasks, and perform several experiments using state-of-the-art R-CNN-based models for datasets in the self-driving domain. Our empirical experiments on a number of datasets reflect that active learning reduces the amount of training data required. We observe that early exploration with instance- rich training sets leads to good performance, and that false positives can have a negative impact if not dealt with appropriately. Furthermore, we perform a qualitative evaluation using autonomous driving data collected from Trondheim, illustrating that active learning can help in selecting more informative images to annotate.

Machine Translation (MT) is an application area of NLP where automatic systems are used to translate text or speech from one language to another while preserving the meaning of the source language. Although there exists a large volume of literature in automatic machine translation of documents in many languages, the translation between English and Tigrigna is less explored. Therefore, we proposed the hybrid approach to address the challenges of applying syntactic reordering rules which align and capture the structural arrangement of words in the source sentence to become more like the target sentences. Two language models were developed- one for English and another for Tigrigna and about 12,000 parallel sentences in four domains and 32,000 bilingual dictionaries were collected for our experiment. The parallel collected corpus was split randomly to 10,800 sentences for training set and 1,200 sentences for testing. Moses open source statistical machine translation system has been used for the experiment to train, tune and decode. The parallel corpus was aligned using the Giza++ toolkit and SRILM was used for building the language model. Three main experiments were conducted using statistical approach, hybrid approach and post-processing technique. According to our experimental result showed good translation output as high as 32.64 BLEU points Google translator and the hybrid approach was found most promising for English-Tigrigna bi-directional translation.

Skull stripping in brain imaging is the removal of the parts of images corresponding to non-brain tissue. Fast and accurate skull stripping is a crucial step for numerous medical brain imaging applications, e.g. registration, segmentation and feature extraction, as it eases subsequent image processing steps. In this work, we propose and compare two novel skull stripping methods based on 2D and 3D convolutional neural networks trained on a large, heterogeneous collection of 2777 clinical 3D T1-weighted MRI images from 1681 healthy subjects. We investigated the performance of the models by testing them on 927 images from 324 subjects set aside from our collection of data, in addition to images from an independent, large brain imaging study: the IXI dataset (n = 556). Our models achieved mean Dice scores higher than 0.978 and Jaccard indices higher than 0.957 on all tests sets, making predictions on new unseen brain MR images in approximately 1.4s for the 3D model and 12.4s for the 2D model. A preliminary exploration of the models’ robustness to variation in the input data showed favourable results when compared to a traditional, well-established skull stripping method. With further research aimed at increasing the models’ robustness, such accurate and fast skull stripping methods can potentially form a useful component of brain MRI analysis pipelines.

Generative adversarial networks (GANs) have seen some success as a way to synthesize training data for supervised machine learning models. In this work, we design two novel approaches for synthetic image generation based on CycleGANs, aimed at generating realistic-looking, class-specific dermoscopic skin lesion images. We evaluate the images’ usefulness as additional training data for a convolutional neural network trained to perform a difficult lesion classification task. We are able to generate visually striking images, but their value for augmenting the classifier’s training data set is low. This is in-line with other researcher’s investigations into similar GAN models, indicating the need for further research into forcing GAN models to produce samples further from the training data distribution, and to find ways of guiding the image generation using feedback from the ultimate classification objective.

The use of Virtual Simulation (VS) for emergency management and Incident Commander (IC) training and assessment has spread during the last decade. In VS, ICs act in computer- simulated 3D incident scenarios, e.g. fire incidents, road traffic collisions etc. Even though VS provides several benefits, there is a history of hesitation to implement and apply it in emergency education. This paper presents the results of a field study performed during the VS training in four classes of IC-students (90 students). The research focus was on the IC students` attitudes and experiences of VS training. Data were collected through observations and post-training questionnaires. The results show that students are predominantly positive towards virtual simulation. 72% of the IC-students state that they experienced presence to the same extent as in live simulation settings, where they experience high presence. Earlier, photorealism was considered to be necessary to provide virtual learning places with high experiences. According to this study, this is not equally important on a general base. The results argue for the benefits of using VS in IC training, even if there are challenges with the implementation. Furthermore, it contributes to a better understanding of user experiences and realism in VS training compared to live simulation.

We discuss several approaches for defining software languages, together with Integrated Development Environments for them. Theoretical foundation is grammar-based models: they can be used where proven correctness of specifications is required. From a practical point of view, we discuss how language specification can be made more accessible by focusing on language workbenches and projectional editing, and discuss how it can be formalized. We also give a brief overview of unconventional ideas to language definition, and outline three open problems connected to the approaches we discuss.