Maximilian V. Harl

We study how the brain efficiently encodes information across the sequential stages of goal-directed behavior — sensing the environment, evaluating options, and selecting actions. Comparing models in which efficient coding and Bayesian decoding act at perception alone, valuation alone, or both, in an experiment that independently varied stimulus-to-value mappings, we find that efficient coding at both stages — each tracking its own prior — best explains behavior. Perception reflects stable environmental structure while value representations adapt rapidly to context, indicating that separable efficient codes at successive stages combine to shape behavior.

As the statistics of sensory environments change, neural systems must adapt to keep their representations useful. Efficient coding prescribes that tuning curves be optimized to the prior, but whether they can adapt rapidly is unclear. We propose a gain-adaptive recurrent network whose gains optimize an efficient-coding objective balancing accuracy against spiking cost; propagating these modulated gains through the network yields quickly-adapting tuning curves. The model captures subtle "adapter-repulsion" effects under peaked priors and predicts fast prior attraction under broader distributions.

Systematic literature reviews (SLRs) are foundational for research but resource-intensive. With the rise of large language models such as ChatGPT, generative AI is increasingly explored to support the SLR process. Following PRISMA 2020, we screened 1,846 publications (Jan 2021–Apr 2025) and analyzed 54 in detail, finding that generative AI is applied most in the screening, search, and writing phases — predominantly via basic prompting with human oversight. Efficiency gains are frequently reported, though several concerns remain.

Business process redesign (BPR) aims to improve business processes but has traditionally been a manual, labor- and time-intensive task. Growing volumes of process data and advances in generative machine learning have opened the door to automated BPR, yet existing approaches run offline on historical data. We argue that performing BPR in runtime and exploiting the predictive power of generative machine learning achieves a higher degree of automation — letting organizations improve their processes proactively — and outline a design-science process for such a technique.

Despite increasing digitization of healthcare providers, no methodology comprehensively describes and aligns the operational, clinical, economic, distributed, and interactive technology surrounding the patient as a "self-practitioner". We examine which of four descriptive concepts — patient flow, patient journey, clinical pathway, and clinical guidelines — best supports comprehensive analytical data modeling and decision support for individualized precision medicine, weighing both patient outcomes and practitioner work-life balance, through a quantitative review of 7,000+ articles from 18 VHB journals and PubMed.

Discovering clinically relevant epistatic interactions (EIs) between single-nucleotide polymorphisms (SNPs) is difficult: existing statistical methods are largely limited to SNP pairs due to the combinatorial explosion of higher-order interactions. We introduce NeEDL (network-based epistasis detection via local search), which uses network medicine to guide EI selection. NeEDL's interactions are an order of magnitude more statistically significant than those of existing tools and comprise, on average, five SNPs; the approach is also amenable to acceleration on quantum computers.

In recent years, large pre-trained deep neural networks (DNNs) have revolutionized the field of computer vision (CV). Although these DNNs have been shown to be very well suited for general image recognition tasks, application in industry is often precluded for three reasons: 1) large pre-trained DNNs are built on hundreds of millions of parameters, making deployment on many devices impossible, 2) the underlying dataset for pre-training consists of general objects, while industrial cases often consist of very specific objects, such as structures on solar wafers, 3) potentially biased pre-trained DNNs raise legal issues for companies. As a remedy, we study neural networks for CV that we train from scratch, using a real-world case from a solar wafer manufacturer. We find that our neural networks achieve similar performances as pre-trained DNNs, even though they consist of far fewer parameters and do not rely on third-party datasets.

We present the Graph Relevance Miner (GRM), a technique based on graph neural networks that computes relevance scores for the activities of a process model with respect to a performance measure. Annotating process models with these scores enables a problem-focused analysis of the business process, directing attention to the activities most worth improving. We validate the predictive quality and the reliability of the generated relevance scores on datasets from several industries.

Predictive business process monitoring (PBPM) forecasts the future behavior of running process instances, yet accurate predictions are often not actionable without explanation. We propose a technique using gated graph neural networks (GGNNs) that makes predictions explainable by visualizing how much each activity in a process contributes to a given prediction — the first use of GGNNs both for PBPM and for rendering such predictions explainable.