We are looking for PhD students who will be part of an interdisciplinary research environment including also international collaborators. Each PhD student is expected to cover the three fronts of theoretical analysis (e.g. performance guarantees), algorithm design and implementation, carrying out research on one or (preferably) two of the following interconnected topics:

• Topic 1: Learning explainable representations capturing optimally the complex spatio-temporal dependencies and geometry (e.g. higher-order, causal dependencies, dynamics in latent spaces, etc.) in multivariate and possibly irregular data, as well as exploiting those learned representations for various inference, reasoning, decision-making and control tasks, considering also robustness and reliability against noise, outliers, data scarcity or missing data. These representations will also allow integrating structural inductive biases or different types of prior knowledge (e.g. domain knowledge, physical laws, ontologies), as well as quantifying and tracking uncertainties, without necessarily distributional or asymptotic assumptions. Both batch and online algorithms will be considered, understanding how to design the best optimization-based training algorithms. This position/topic will be within the SIGIPRO Department at SimulaMet, with possible collaborations across other Departments at Simula, as well as other national and international partners in SURE-AI. 
• Topic 2: Constrained machine learning algorithms for training, inference and control, where constraints may come from different perspectives, such as safety, risk-aware constraints, physical systems principles, ethical constraints, fairness, privacy, regulations, or constraints in terms of some of the rewards for instance in the context of reinforcement learning. This area is relevant for many problems, including adapting and fine-tuning foundation models under different constraints or human alignment requirements, or incorporating various risk-aware measures in decision-making or control problems (e.g. reinforcement learning), or ethically embedded machine learning algorithms. This position/topic will be within the SIGIPRO Department at SimulaMet, with possible collaborations across other Departments at Simula, as well as other national and international partners in SURE-AI. 
• Topic 3: Efficient methods and computational architectures for scalable, resource and data efficient, and domain-adaptable machine learning, considering both centralized and decentralized (multi-agent inter-connected systems) machine learning scenarios. This includes co-design of training and inference algorithms jointly with advanced high-performance disaggregated computation architectures, as well as alternative disruptive architectures, such as (most importantly) neuromorphic computing or (possibly) quantum computing. This may involve algorithmic reformulations that increase arithmetic intensity, data locality, memory efficient attention mechanisms and joint computation-communication efficient strategies, surpassing the “memory wall” bottleneck and reducing energy demands. In the context of large models, such as foundation models (e.g. LLMs), this may involve techniques such as low-rank and sparsity approximations, tensor decomposition methods, model pruning, quantization, compression, splitting methods, stochastic and dynamical systems approximations (e.g. neural ODEs/SDEs, neural operators), and approximate mixed-precision training and inference strategies. This is a multidisciplinary topic and will include collaboration across several departments in Simula, including HOST, SCAN, as well as an additional postdoctoral researcher from SIGIPRO through the RCN Program “Recruitment of Talented Researchers to Norway” for National Centres.
• Topic 4: Generalizable, continual and online learning methods designed to operate under strict computational, data, and reliability constraints. By shifting from offline, centralized training to continual and online learning, this topic addresses time-varying data distributions, adaptation across tasks or environments, and non-IID data scenarios where traditional benchmarks and strategies typically fail. Generalization in inference and control problems can be increased by leveraging structural dependence representation learning and counterfactual reasoning to enable out-of-distribution generalization and active planning, ensuring robustness against noise and unseen samples. Automated hyperparameter optimization is also relevant to adapt to non-stationary distribution shifts as well as plasticity to adapt reliably to new tasks and environments, and maintain stability. In addition to the design of algorithms, it is important also to characterize the generalization capability in terms of providing guarantees, generalization error bounds, or in cases of transfer learning across tasks, characterizing the propagation of uncertainties. This position/topic will be within the SIGIPRO Department at SimulaMet, with possible collaborations across other Departments at Simula, as well as other national and international partners in SURE-AI. 

All these projects/topics will incorporate real data in the context of impactful applications within the verticals of energy and environment, finance and societal systems, in cooperation with some of the public or private partners involved and postdoctoral/senior researchers in the SURE-AI Centre.