A5060969026- Sparse and Compressive Sensing Techniques
- Medical Imaging Techniques and Applications
- Image and Signal Denoising Methods
- Mathematical Analysis and Transform Methods
- Machine Learning and Algorithms
- Model Reduction and Neural Networks
- Advanced MRI Techniques and Applications
- Photoacoustic and Ultrasonic Imaging
- Advanced X-ray and CT Imaging
- Stochastic Gradient Optimization Techniques
- Adversarial Robustness in Machine Learning
- Artificial Intelligence in Healthcare and Education
- Digital Filter Design and Implementation
- Machine Learning and ELM
- Advanced Computational Techniques and Applications
- Cardiac Arrest and Resuscitation
- Disaster Response and Management
- Probabilistic and Robust Engineering Design
- Anomaly Detection Techniques and Applications
- Advanced Scientific Research Methods
- Computability, Logic, AI Algorithms
- Machine Learning and Data Classification
- Control Systems and Identification
- Neural Networks and Applications
- Ethics and Social Impacts of AI
University of Oslo
2020-2025
Deep learning, due to its unprecedented success in tasks such as image classification, has emerged a new tool reconstruction with potential change the field. In this paper, we demonstrate crucial phenomenon: learning typically yields unstable methods for reconstruction. The instabilities usually occur several forms: 1) Certain tiny, almost undetectable perturbations, both and sampling domain, may result severe artefacts reconstruction; 2) small structural change, example, tumor, not be...
Significance Instability is the Achilles’ heel of modern artificial intelligence (AI) and a paradox, with training algorithms finding unstable neural networks (NNs) despite existence stable ones. This foundational issue relates to Smale’s 18th mathematical problem for 21st century on limits AI. By expanding methodologies initiated by Gödel Turing, we demonstrate limitations (even randomized) computing NNs. Despite numerous results NNs great approximation properties, only in specific cases do...
Abstract We revisit the construction of wavelets on interval with various degrees polynomial exactness, and explain how existing schemes for orthogonal- Spline can be extended to compactly supported delay-normalized wavelets. The contribution differs substantially from previous ones in results are stated deduced: linear algebra notation is exploited more heavily, use sums complicated index reduced. This eases translation software, a general open source implementation, which uses deductions...
Artificial Intelligence (AI) has the potential to greatly improve delivery of healthcare and other services that advance population health wellbeing. However, use AI in also brings risks may cause unintended harm. To guide future developments AI, High-Level Expert Group on set up by European Commission (EC), recently published ethics guidelines for what it terms “trustworthy” AI. These are aimed at a variety stakeholders, especially guiding practitioners toward more ethical robust...
Infinite-dimensional compressed sensing deals with the recovery of analog signals (functions) from linear measurements, often in form integral transforms such as Fourier transform. This framework is well-suited to many real-world inverse problems, which are typically modelled infinite-dimensional spaces, and where application finite-dimensional approaches can lead noticeable artefacts. Another typical feature problems that not only sparse some dictionary, but possess a so-called local...
There are two big unsolved mathematical questions in artificial intelligence (AI): (1) Why is deep learning so successful classification problems and (2) why neural nets based on at the same time universally unstable, where instabilities make networks vulnerable to adversarial attacks. We present a solution these that can be summed up words; false structures. Indeed, does not learn original structures humans use when recognising images (cats have whiskers, paws, fur, pointy ears, etc), but...
Methods inspired by Artificial Intelligence (AI) are starting to fundamentally change computational science and engineering through breakthrough performances on challenging problems. However, reliability trustworthiness of such techniques is becoming a major concern. In inverse problems in imaging, the focus this paper, there increasing empirical evidence that methods may suffer from hallucinations, i.e., false, but realistic-looking artifacts; instability, sensitivity perturbations data;...
Most datasets used for supervised machine learning consist of a single label per data point. However, in cases where more information than just the class is available, would it be possible to train models efficiently? We introduce two novel model architectures, which we call hybrid concept-based models, that using both labels and additional dataset referred as concepts. In order thoroughly assess their performance, ConceptShapes, an open flexible with concept labels. show outperform standard...
Understanding the implicit regularization imposed by neural network architectures and gradient based optimization methods is a key challenge in deep learning AI. In this work we provide sharp results for flow of Diagonal Linear Networks (DLNs) over-parameterized regression setting and, potentially surprisingly, link to phenomenon phase transitions generalized hardness approximation (GHA). GHA generalizes from computer science to, among others, continuous robust optimization. It well-known...
Undersampled inverse problems occur everywhere in the sciences including medical imaging, radar, astronomy etc., yielding underdetermined linear or non-linear reconstruction problems. There are now a myriad of techniques to design decoders that can tackle such problems, ranging from optimization based approaches, as compressed sensing, deep learning (DL), and variants between two techniques. The variety methods begs for unifying approach determine existence optimal fundamental accuracy...
Recovering a signal (function) from finitely many binary or Fourier samples is one of the core problems in modern imaging, and by now there exist plethora methods for recovering such samples. Examples which can utilize wavelet reconstruction include generalized sampling, infinite-dimensional compressive sensing, parameterized-background data-weak (PBDW) method, etc. However, any these to be applied practice, accurate fast modeling an $N \times M$ section change-of-basis matrix between...
Instability is AI's Achilles’ heel. We show the following paradox: there are cases where stable and accurate AI exists, but it can never be trained by any algorithm. initiate a foundations theory for when - such programme will shape political legal decision-making in coming decades, have significant impact on markets technologies.
Recovering a signal (function) from finitely many binary or Fourier samples is one of the core problems in modern medical imaging, and by now there exist plethora methods for recovering such samples. Examples methods, which can utilise wavelet reconstruction, include generalised sampling, infinite-dimensional compressive sensing, parameterised-background data-weak (PBDW) method etc. However, any these to be applied practice, accurate fast modelling an $N \times M$ section change-of-basis...