ELLIS Unit

Supervisors

Name: Serge Thill / Roel Vertegaal
Email: serge.thill [at] donders.ru.nl (serge[dot]thill[at]donders[dot]ru[dot]nl) and roel.vertegaal [at] ru.nl (roel[dot]vertegaal[at]ru[dot]nl) 
Website: Department: Human-Centered Interactive Systems / Human Media Lab
Faculty: Faculty of Social Sciences and Faculty of Sciences

Project Description

Background/Motivation:

Recent work (Vertegaal and Merrit; 2024) describes a framework, derived from Bayesian inference, for analysing both performance and error in Human-Computer Interaction tasks that can be used to estimate mental load without the use of a secondary task. This is of interest in various contexts of human-machine teaming where the mental load of a human team member is a relevant factor in determining the behaviour of machines such as robots in a collaborative task (see Carissoli et al, 2023, for a recent review). Typically, one would use a secondary task to measure cognitive load experimentally, such as counting audible beeps (where mental load is taken to be correlated with the frequency of missing beeps), but this is not a viable solution in an ecologically valid setting. A challenge with the proposed framework, meanwhile, is that it requires a good definition of the primary task that the user is carrying out. In realistic human-robot interactions, this may be difficult to define since it is not necessarily known a priori what humans will take into account. 

Overall, there is both a need to develop the model further for ecologically valid examples of human-robot interactions, validate its functioning, and demonstrate the degree to which it can help a robot adjust its behaviour based on workload. 

Research Question/Goal:

Can mental load be estimated in human-robot collaborative tasks in the same way this can be done in HCI tasks, and does it matter for making the robot adapt to the user?

Method:

• Extend the existing framework to tasks that are collaborative in nature
• Validate this in a study that compares mental load predictions based on the framework with those coming from a secondary task
• Define and implement a collaborative task between a human and a robot (this could be done in simulation) where the mental load of the user is likely to fluctuate, and contrast task performance between an adaptive robot that can adjust based on estimated mental workload and one that cannot. Example tasks can be inspired by games that use mental load as a game (e.g. the game Overcooked)

Student requirements

Experience with / interest in: Bayesian inference, HRI/HCI, sufficient programming skills for developing a controller for a robot (or simulated agent) and task, mental load modelling

Type of project: 30 EC Research project

Expected time frame: 6 Months

Literature 

1. Carissoli, C., Negri, L., Bassi, M., Storm, F. A., & Delle Fave, A. (2023). Mental Workload and Human-Robot Interaction in Collaborative Tasks: A Scoping Review. International Journal of Human–Computer Interaction, 1–20. https://doi.org/10.1080/10447318.2023.2254639

2.Vertegaal, R. and Merritt, T. Theories for Human-Computer Interaction. Departmental Report, iCIS, Radboud University, 2024.

3. Friston, K, Kilner, J and Harrison, L. 2006. A free energy principle for the brain. Journal of physiology-Paris 100, 1-3 (2006), 70–87.

Supervisors

Name: Maris Galesloot MSc & Prof. Dr. Nils Jansen
Email: maris.galesloot [at] ru.nl (maris[dot]galesloot[at]ru[dot]nl), nils.jansen [at] ru.nl (nils[dot]jansen[at]ru[dot]nl) 
Website: https://ai-fm.org/
Department: Department of Software Science @ iCIS
Faculty: FNWI

Project Description

Background/Motivation:

Computing optimal sequential decision-making policies for autonomous agents in partially observable environments, for instance, where the underlying state is not directly observable, is generally a hard problem. Recent work in reinforcement learning (RL) employing recurrent neural networks (RNNs) as a policy network has seen great success in such environments (Ni et al., 2022). RNNs learn approximate representations of the history of the system to enable memory-based policies. The policies are trained using RL algorithms on a dataset of transitions, which are collected by simulating steps in the model of the environment. However, by neglecting the information in the model, RL algorithms generally require an enormous number of simulation steps. Furthermore, the computed policies lack guarantees on their performance and are not interpretable.

Research Question/Goal:

Employing RNNs in a model-based approach gives rise to the question of how to design an efficient optimization objective for the RNN policy. Prior work is able to exploit the information from the model to speed up learning (Carr et al., 2021). However, such model-based methods have so far relied on sub-optimal learning targets, which are computed directly from the model, instead of the usual optimization objective in RL: maximizing the expected rewards.

Our proposal is inspired by the success of AlphaZero (Silver et al., 2017), a Monte Carlo tree search (MCTS) method equipped with neural networks to iterate over the space of policies. Such an approach can find solutions to difficult problems with a fully observable state (such as Go and Chess) but is understudied in partially observable environments. We aim to bridge this gap.

Method:

The idea is to combine POMCP (Silver and Veness, 2010), a well-known variant of the MCTS algorithm suited for partially observable environments, with an RNN architecture to find a recurrent neural policy and associated memory representation. Then, we use an existing technique to discretize the memory representation (Koul et al., 2019). This discretization of the RNN yields a finite-state machine as a policy, for which we can exactly evaluate the performance on the model.

After implementing the sketched framework, which by itself is a major contribution, there are many things to explore. For instance, we can consider how to exploit the information gained from computing the policy's performance on the model to steer further search iterations.

Student requirements

Experience with / interest in: machine learning, reinforcement learning, POMDPs 
Type of project (master thesis, internship): Both are possible.
Expected time frame: 6 months

Literature

1. Carr, S., Jansen, N., and Topcu, U. (2021). Task-aware verifiable rnn-based policies for partially observable markov decision processes. J. Artif. Intell. Res., 72:819–847.
2. Koul, A., Fern, A., and Greydanus, S. (2019). Learning finite state representations of recurrent policy networks. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net.
3. Ni, T., Eysenbach, B., and Salakhutdinov, R. (2022). Recurrent model-free RL can be a strong baseline for many pomdps. In ICML, volume 162 of Proceedings of Machine Learning Research, pages 16691–16723. PMLR.
4. Silver, D., Schrittwieser, J., Simonyan, K., Antonoglou, I., Huang, A., Guez, A., Hubert, T., Baker, L., Lai, M., Bolton, A., Chen, Y., Lillicrap, T. P., Hui, F., Sifre, L., van den Driessche, G., Graepel, T., and Hassabis, D. (2017). Mastering the game of go without human knowledge. Nat., 550(7676):354–359.
5. Silver, D. and Veness, J. (2010). Monte-carlo planning in large POMDPs. In NIPS, pages 2164–2172. Curran Associates,Inc.

Supervisors 

Name: Maris Galesloot MSc, Prof. Dr. Nils Jansen
Email: maris.galesloot [at] ru.nl
Website: ai-fm.org
Department: Software Science @ iCIS (CS department)
Faculty: FNWI

Project Description

Background/Motivation:

Reinforcement learning (RL) is the prevalent framework for the control of agents in high-dimensional environments. In realistic domains, such environments are partially observable, for instance, because the agent needs to act based on vision or sensory inputs or requires reasoning over past actions [1]. It is well-known that agents require memory to act optimally in partially observable environments. 

Research Question/Goal:

In deep RL, various architectures have been proposed to enable agents with memory, such as recurrent neural networks [2]. However, there is still much progress to be made in (1) how to balance the short- and long-term memory of the agent, (2) how to accurately represent the history of observations of the agent's environment, and (3) how to learn efficiently with memory.

Method:

We will test new memory architectures for (deep) RL algorithms, such as the newly introduced xLSTM [3], and compare it to existing architectures and/or look into developing a (specialised) architecture to accommodate memory for the agent.

Student requirements

Experience with / interest in: reinforcement learning, machine learning, deep learning

Type of project (master thesis, internship (only for companies)): can be both

Expected time frame: 6 months

Literature 

1. Recurrent model-free RL can be a strong baseline for many POMDPs. ICML 2022. Ni, Eysenbach, and Salakhutdinov.
2. Learning belief representations for partially observable deep RL. ICML 2023. Wang, Li, Klassen, Icarte, and McIlraith. 
3. xLSTM: Extended Long Short-Term Memory. arXiv 2024. Maximilian Beck, Korbinian Pöppel, Markus Spanring, Andreas Auer, Oleksandra Prudnikova, Michael Kopp, Günter Klambauer, Johannes Brandstetter, Sepp Hochreiter.

Supervisors 

Name: Maris Galesloot MSc, Dr. Thiago D. Simão, Prof. Dr. Nils Jansen
Email: maris.galesloot [at] ru.nl
Website: ai-fm.org
Department: Software Science @ iCIS (CS department)
Faculty: FNWI

Project Description

Background/Motivation:

Many modern problems, such as the optimisation of cellular networks, can be formulated as a multi-agent system with cost constraints [1]. Using this formulation, we can use machine learning methods, such as multi-agent reinforcement learning, to find the optimal performance for the agents. Because of the rapid increase of complexity when many agents are involved, we often need to distribute the optimisation problem in a graphical structure that models (sparse) interactions between the agents. 

Research Question/Goal:

Existing techniques exist to optimise an unconstrained multi-agent problem given a graph [2]. However, it remains an open problem to efficiently formulate and optimise in a distributed manner while satisfying constraints.

Method:

You will work on (multi-agent) reinforcement learning algorithms, such as Monte Carlo tree search, under cost constraints [3]. In particular, we are interested in determining if we can also decompose the optimisation objective using the graph, given that we must satisfy a maximum budget of expected costs. You can analyse the problem theoretically and/or extend the aforementioned (existing) algorithms. In the latter case, your supervisors can supply you with existing code from which you can start working.

Student requirements

Experience with / interest in: reinforcement learning, multi-agent systems, machine learning
Type of project (master thesis, internship (only for companies)): can be both
Expected time frame: 6 months

Literature 

1. Albin Larsson Forsberg, Alexandros Nikou, Aneta Vulgarakis Feljan, Jana Tumova. Network Parameter Control in Cellular Networks through Graph-Based Multi-Agent Constrained Reinforcement Learning. IEEE, 2023.
2. Maris Galesloot, Thiago D. Simão, Sebastian Junges, Nils Jansen. Factored Online Planning in Many-Agent POMDPs. AAAI, 2024.
3. Jongmin Lee, Geon-hyeong Kim, Pascal Poupart, Kee-Eung Kim. Monte-Carlo Tree Search for Constrained POMDPs. NeurIPS, 2018.
   
    

Supervisors

Name: Sarah de Boer, Msc
Email: sarah.deboer [at] radboudumc.nl
Website: https://www.diagnijmegen.nl/people/sarah-de-boer/
Department: Beeldvorming, Diagnostic Image Analysis Group
Faculty: Radboudumc

Name: Alessa Hering, PhD
Email: alessa.hering [at] radboudumc.nl
Website: https://www.diagnijmegen.nl/people/alessa-hering/
Department: Beeldvorming, Diagnostic Image Analysis Group
Faculty: Radboudumc

Project Description

Background/Motivation:

The COMFORT project (Computational Models FOR patienT stratification in urologic cancers – Creating robust and trustworthy multimodal AI for health care) (https://comfort-ai.eu) aims to develop robust and trustworthy multimodal AI systems to enhance clinical outcome for prostate and kidney cancer patients. Our goal is to create internationally and interdisciplinarily validated decision support systems that improve clinical prognosis, patient stratification and individualized therapy options. Currently, our team is developing an AI-based pipeline for segmenting and classifying renal masses (also called lesions) in computer tomography (CT) scans. The pipeline uses the segmentation masks as detection guidance to find the lesions in the CT scans, which has shown promising results. However, it also presents challenges with false positives and negatives leading to either missed lesions or incorrect lesion identifications. To address this issue, we are interested in adding a detection algorithm to this pipeline to refine our system’s accuracy in locating lesions. This algorithm would ultimately together with the segmentation algorithm locate the lesions and find their volumes/shapes. Given the limited availability of annotated data, we are particularly interested in investigating unsupervised learning approaches. Techniques such as diffusion models, Variational Autoencoders (VAEs), Generative Adversarial Networks (GANs), and inpainting methods show potential for enhancing the lesion detection capabilities without extensive labeled datasets. 

Research Question/Goal: 

Can the integration of unsupervised learning methods such as diffusion models, VAEs, GANs, and inpainting improve the detection of kidney lesions in CT scans?

Method:

In medical image analysis, nnUNet represents the state-of-the-art algorithm (Isensee et al., 2021) for segmentation tasks. It employes a self-configering mechanism that adjusts the Unet architecture (Ronneberger et al., 2015) and various hyperparameters automatically to suit specific characteristics of the training dataset. Complementing this, the nnDetection algorithm (Baumgartner et al., 2021) extends nnUNet principles to tackle object detection in medical imaging.

Although, this framework provides a good starting point for this project, the supervised nature of this algorithm requires it to have annotated data for training. There are some publicly available kidney cancer datasets (Heller et al., 2021; Humpire-Mamani et al., 2023), however they are annotated for segmentation purposes. Consequently, while the nnDetection framework provides a valuable baseline, the need for annotated data presents a significant limitation. Therefore, the aim of this project is to investigate unsupervised learning approaches and data synthesis techniques to overcome these limitations. Chen et al. (2020) developed a method that uses a prior learned on images of healthy patients. The model removes abnormal lesions, which are not represented in the prior, from the image during the restoration step, after which the detection is formed by the difference between the original image and the restored image. By exploring methods such as those proposed by (Chen et al., 2020), this project seeks to develop detection algorithms that do not rely on extensive annotated datasets. Furthermore, advanced data synthesis techniques (Koetzier et al., 2024), or inpainting are potentially interesting approaches to augment our training dataset artificially. These approaches generate realistic imaging data and providing additional resources for training detection models.

Student requirements

Experience with / interest in: AI, deep learning, medical image analysis
Type of project (master thesis, internship (only for companies)): Master thesis
Expected time frame: 6 months

Literature 

1. https://comfort-ai.eu
2. Baumgartner, M., Jäger, P. F., Isensee, F., & Maier-Hein, K. H. (2021). nnDetection: A Selfconfiguring Method for Medical Object Detection. In M. de Bruijne, P. C. Cattin, S. Cotin, N. Padoy, S. Speidel, Y. Zheng, & C. Essert (Eds.), Medical Image Computing and Computer Assisted Intervention – MICCAI 2021 (pp. 530–539). Springer International Publishing. https://doi.org/10.1007/978-3-030-87240-3_51
3. Heller, N., Isensee, F., Maier-Hein, K. H., Hou, X., Xie, C., Li, F., Nan, Y., Mu, G., Lin, Z., Han, M., Yao, G., Gao, Y., Zhang, Y., Wang, Y., Hou, F., Yang, J., Xiong, G., Tian, J., Zhong, C., … Weight, C. (2021). The state of the art in kidney and kidney tumor segmentation in contrast-enhanced CT imaging: Results of the KiTS19 challenge. Medical Image Analysis, 67, 101821. https://doi.org/10.1016/j.media.2020.101821
4. Koetzier, L. R., Wu, J., Mastrodicasa, D., Lutz, A., Chung, M., Koszek, W. A., Pratap, J., Chaudhari, A. S., Rajpurkar, P., Lungren, M. P., & Willemink, M. J. (2024). Generating Synthetic Data for Medical Imaging. Radiology, 312(3), e232471. https://doi.org/10.1148/radiol.232471
5. Chen, X., You, S., Tezcan, K. C., & Konukoglu, E. (2020). Unsupervised lesion detection via image restoration with a normative prior. Medical Image Analysis, 64, 101713. https://doi.org/10.1016/j.media.2020.101713

Supervisors

Name: Yuzhen Qin, Marcel van Gerven
Email: yuzhen.qin [at] donders.ru.nl
Website: https://yuzhenqin90.github.io/, https://www.ru.nl/personen/gerven-m-van
Department: Machine Learning and Natural Computing
Faculty: Social Science

Project Description

Motivation: 

Path planning is vital for robotics, enabling robots to navigate from a starting point to a goal while avoiding obstacles. It is crucial for autonomous systems like self-driving cars, drones, and warehouse automation. However, in dynamic environments, path planning is challenging, as real-time decisions require frequent replanning to account for moving obstacles and shifting goals, demanding significant computational resources. Current algorithms, especially in high-dimensional spaces, are computationally expensive. This project proposes a novel solution using neuromorphic computing based on physically coupled oscillators. These oscillatory networks naturally settle into minimum-energy states, a property leveraged to solve complex NP-hard optimization problems. Hardware implementations have shown potential to solve such problems within miniseconds or even microseconds. Applying this approach to path planning could provide ultrafast, energy-efficient solutions, revolutionizing real-time navigation in dynamic environments.

Research Question/Goal: 

This project explores a simplified path planning scenario, where a robot navigates through a grid world with moving obstacles. The robot must frequently re-calculate the optimal path to reach a potentially moving target. Our primary research question is: Can oscillator-based neuromorphic computing be used to dynamically plan paths quickly and efficiently in this environment? Answering this question will provide proof-of-concept results, demonstrating the potential for real-world applications in robot path planning.

Methods: 

1. We will utilize dynamical systems theory and graph theory to reformulate the path planning problem as an oscillatory network. This network will inherently solve the combinatorial optimization challenges of path planning. 
2. We will conduct simulations of oscillatory networks using our theoretical framework to assess their effectiveness in solving dynamic path planning problems. Successful validation will demonstrate the feasibility of using neuromorphic computing for real-time path planning in robotics.

Student requirements

Experience with / interest in: Theoretical & algorithmic machine learning, dynamical systems theory
Type of project: master thesis
Expected time frame: 01-02-2025 to 31-07-2025

Literature

1. Wang, Tianshi, et al. "Solving combinatorial optimisation problems using oscillator based Ising machines." Natural Computing 20.2 (2021): 287-306.
2. Mohseni, Naeimeh, Peter L. McMahon, and Tim Byrnes. "Ising machines as hardware solvers of combinatorial optimization problems." Nature Reviews Physics 4.6 (2022): 363-379.

Supervisors

Name: Marzieh Hassanshahi Varposhti
Email: Marzieh.hassanshahi [at] donders.ru.nl (Marzieh[dot]hassanshahi[at]donders[dot]ru[dot]nl) 
Department: Machine learning and neural computation
Faculty: Social Science

Name: Mahyar Shahsavari 
Email: mahyar.shahsavari [at] donders.ru.nl
Website: https://www.ru.nl/en/people/shahsavari-m 
Department: Machine learning and neural computation
Faculty: Social Science

Project Description

Background/Motivation:

State-space models like Mamba are gaining attention as alternatives to traditional Transformer models due to their ability to handle long-range dependencies and temporal dynamics more efficiently. While Transformers rely on self-attention mechanisms, which can become computationally expensive, especially when scaling to longer sequences, state-space models offer a more computationally tractable approach. Mamba, in particular, leverages continuous representations that allow for more efficient temporal information processing with fewer computational resources.

One of the primary reasons to consider using state-space models over Transformers is their inherent ability to capture long-term dependencies while reducing the complexity associated with processing extensive sequences. This advantage becomes particularly pronounced when adapting the Mamba model into a neuromorphic format. In neuromorphic computing, where energy efficiency is a critical consideration, spiking neural networks (SNNs) offer a significant reduction in power consumption compared to traditional models.

A neuromorphic version of Mamba would further enhance the model’s ability to govern temporal information by capitalizing on the event-driven nature of SNNs. In SNNs, neurons only fire when there is a change in input, making them well-suited for handling temporal data, especially when sparsity is a factor. This characteristic aligns naturally with the temporal dynamics Mamba is designed to manage, allowing for a more efficient and biologically plausible way of processing sequential information.

The most significant advantage of adopting a spiking version of Mamba is the energy efficiency it brings. In hardware implementations, particularly in edge computing or resource-constrained environments, minimizing energy consumption is crucial. Spiking networks, due to their event-driven nature, use less power by firing only when necessary, which could result in a much more efficient model compared to traditional, continuously active neural architectures like Transformers.

Beyond energy efficiency, a spiking Mamba model could offer improved scalability for real-time applications that require quick responses to dynamic, temporal patterns, such as robotics, autonomous systems, and other hardware-based AI systems. By leveraging the temporal efficiency of spiking neurons, such systems could better handle real-world, time-varying signals with reduced latency and higher throughput.

In summary, by transitioning to a spiking version of the Mamba state-space model, there is the potential for a more energy-efficient, hardware-friendly system that better captures temporal
dependencies. This not only makes it a suitable alternative to Transformers but also positions it as a promising approach for neuromorphic applications where efficiency, scalability, and real-time performance are paramount.

Research Question/Goal:

The goal of this project is to integrate STDP, a biologically inspired learning rule that adjusts synaptic weights based on the precise timing of spikes, into the neuromorphic version of the Mamba model. This modification aims to improve the model’s capacity to capture and process temporal dependencies in data. By leveraging STDP, the project seeks to make the Mamba model more adaptive to dynamic patterns and sequences, optimizing its performance in tasks requiring fine-grained temporal sensitivity. Ultimately, this research could lead to more efficient and biologically plausible models that operate effectively in neuromorphic hardware, with potential applications in time-sensitive domains such as robotics and real-time signal processing.

Method:

1. Literature Review: Conduct an extensive review of existing state-space models, MAMBA applications, and neuromorphic computing methods.
2. Model Development: Develop neuromorphic versions of state-space models, drawing from existing neuromorphic computing frameworks such as spiking neural networks (SNNs).
3. Verification with MAMBA: Apply MAMBA to verify the correctness of the developed neuromorphic models, ensuring they meet the required specifications.
4. Experimental Comparison: Compare the performance of traditional and neuromorphic state-space models in terms of speed, energy efficiency, and scalability.
5. Analysis and Reporting: Present a detailed analysis of the results, discussing the potential of neuromorphic state-space models to replace conventional implementations.

Student requirements

Experience with / interest in:

• Proficient in Python programming and artificial intelligence (AI).
• Experience with at least one machine learning framework (e.g., TensorFlow, PyTorch, or similar).
• Strong interest in spiking neural networks (SNNs) and neuromorphic computing.
• Familiarity with state-space models or transformers is a plus but not required.
• Ability to work independently and collaboratively in a research setting.

Type of project: Master thesis
Expected time frame: 6 months

Literature 

1. Efficiently modeling long sequences with structured state spaces
2. Mamba: Linear-time sequence modeling with selective state spaces.
3. Q-S5: Towards Quantized State Space Models
4. Spike-driven Transformer
5. TE-Spikformer:Temporal-enhanced spiking neural network with transformer

Supervisors 

Name: Max Hinne           
Email: max.hinne [at] donders.ru.nl (max[dot]hinne[at]donders[dot]ru[dot]nl)            
Website: https://maxhinne.github.io/uncertainty-in-complex-systems/                
Department: Artificial Intelligence
Faculty: Social Science   

Project Description

Background/Motivation: 

Bayesian hierarchical modelling (BHM) is a powerful statistical tool. It combines the hierarchical structure that is inherent in many real-world datasets, and this enables it to:

1. improve statistical strength
2. identify heterogeneity in datasets
3. quantify uncertainty, and consequently
4. improve inferences and predictions compared to flat models [1]. 

Determining the optimal structure and distributional assumptions of a BHM is a major challenge. In many domains, individual differences and unobserved covariates can cause the actual distribution to contain complicated structure, for which no one-size-fits-all solution exists (see [2, 3] for some examples in neuroscience and behavioural science). For specific use cases, statisticians together with domain experts can derive an appropriate model that takes all this into consideration, but this is time consuming and hard to scale. This challenge is amplified when data are scarce, such as in modelling rare diseases, or when data are observed sequentially.

Research Question/Goal: 

How can we automate the creation of the structure and distributional assumptions of a Bayesian hierarchical model in an online setting?

Method: 

The key to this project is to consider the BHM as a random variable itself, and estimate it from observations. However, this random variable is obviously very high-dimensional (it covers all possible Bayesian hierarchical models of the domain), so care must be taken to make the approach scalable. We use approximate Bayesian inference techniques such as Laplace approximations, Bayesian quadrature, and Sequential Monte Carlo [4] to evaluate the fit of candidate models.

On the experimental side of this project, we apply the prototype of the new automated BHM approach to datasets that have also been modelled with handmade hierarchical models, and see how our solutions compare to the bespoke ones in terms of inferences, predictions, and interpretability. 

Ultimately, this project will provide a first step towards a fully automated Bayesian hierarchical statistician.

Student requirements

Experience with / interest in: Bayesian modelling, machine learning, Bayesian fundamentals               

Type of project: Master thesis   

Expected time frame: This project is most suited for a 45 ECTS thesis, of which 6 months are funded by the ELLIS Unit Nijmegen.

Literature

1. Veenman, M., Stefan, A. M., & Haaf, J. M. (2023). Bayesian hierarchical modeling: An introduction and reassessment. Behavior Research Methods, 56(5), 4600–4631. https://doi.org/10.3758/s13428-023-02204-3 
2. De Boer, A. A. A., Bayer, J. M. M., Kia, S. M., Rutherford, S., Zabihi, M., Fraza, C., Barkema, P., Westlye, L. T., Andreassen, O. A., Hinne, M., Beckmann, C. F., & Marquand, A. (2024). Non-Gaussian normative modelling with hierarchical Bayesian regression. Imaging Neuroscience, 2, 1–36. https://doi.org/10.1162/imag_a_00132
3. Dijkstra, S. H. E., Hinne, M., Segers, E., & Molenaar, I. (2023). Clustering children’s learning behaviour to identify self-regulated learning support needs. Computers in Human Behavior, 145, 107754. https://doi.org/10.1016/j.chb.2023.107754
4. Zhou, Y., Johansen, A. M., & Aston, J. A. D. (2016). Toward Automatic Model Comparison: An Adaptive Sequential Monte Carlo Approach. Journal of Computational and Graphical Statistics, 25(3), 701–726. https://doi.org/10.1080/10618600.2015.1060885

Supervisors 

Name: Silvan Quax, Francesco Ciompi, Vincent Stirler
Email: silvan.quax [at] radboudumc.nl
Department: Radiology, Surgery
Faculty: Medical Sciences

Project Description

Background/Motivation:

Initial trauma assessment of severely injured patients is a fast-paced and challenging activity that is unlike most other clinical activities. A multidisciplinary team of varying composition is responsible for primary survey and resuscitation.[1] It performs multiple evaluations and procedures simultaneously and will focus on teamwork and effective communication day and night. The timely diagnosis of injuries is mandatory to initiate appropriate treatments and to prevent further harm at a later point in time.

Many critical decisions are made in during the initial trauma assessment. Fitzgerald et al reported that a critical decision linked to a lifesaving intervention was made every 72 seconds in the first 30 minutes of the primary survey.[2] However, diagnostic errors are more likely to occur in the emergency department than anywhere else.[3] It is reported that "3.2 errors were made per patient in the first 30 minutes of the primary survey and resuscitation."[2] Besides cognitive bias of the team members, typical circumstances are relevant that mitigate diagnostic performance.[4,5] Relevant clinical information is incomplete or absent at presentation. Mental and physical performance impacted by fatigue from working during duty hours. Excessive workloads shorten the time spent on examinations and may lead to hasty assessments. Distractions, such as telephone calls, are repeatedly present and require one to toggle with medical tasks. It is easy to imagine that the initial trauma assessment is prone to error.

Abdominal trauma is present in 7-10% of all trauma victims, with the liver and the spleen being the most common injured organs with a prevalence of 36% and 32% respectively. Their relatively large size, position in the intraperitoneal cavity, and abundant vascular supply make them susceptible to damage and potential sources of catastrophic bleeding. 

Especially the detection of vascular injuries is critical in the initial assessment of a trauma patient, since it is imperative to treat such injuries as soon as possible. The Radboudumc has a unique data set of vascular injuries available as one of the largest trauma centers of the Netherlands. Developing an accurate AI algorithm to detect these traumas will make a significant impact by improving our patient care.

Research Question/Goal:

A deep learning-based model should be developed for the automatic segmentation of vascular injuries on abdominal CT images. Subsequent goals are the classification of the severity of the injury, and the prediction of the probability of rebleeding within the first couple of days following trauma.

Method:

A U-Net based deep learning segmentation algorithm will be the starting point for this project and will be compared to other deep learning based segmentation methods. For the subsequent goals, several classification architectures will be evaluated and compared to the accuracy of medical professionals.

Student requirements

Experience with / interest in: Deep Learning, Python, Pytorch, Healthcare applications
Type of project: Master thesis
Expected time frame: 6 months

Literature

1. Advanced Trauma Life Support (ATLS): The Ninth Edition. J Trauma Acute Care Surg. 2013;74(5)
2. Fitzgerald et al. Trauma resuscitation errors and computer-assisted decision support. Arch Surg. 2011;146(2):218-225.
3. Sevdalis et al. Diagnostic error in a national incident reporting system in the UK.  J Eval Clin Pract. 2010;16(6):1276-1281
4. Leape. Error in Medicine. JAMA J Am Med Assoc. 1994;272(23):1851
5. Waite et al. Interpretive error in radiology. Am J Roentgenol. 2017;208(4):739-749

Supervisors 

Name: Martha Larson    
Email: martha.larson [at] ru.nl
Website: https://www.ru.nl/en/people/larson-m 
Department: iCIS
Faculty: FNWI

Project Description

Background/Motivation:

Recital 26 of the GDPR states that “to determine whether a natural person is identifiable, account should be taken of all the means reasonably likely to be used”. In this project, we apply deep learning approaches to determine whether or not natural persons are identifiable as the “means” considered grows larger and closer to what is “reasonably likely to be used”. Starting from spoken audio that has been anonymized using state-of-the-art methods (adversarial learning & disentanglement), you will gradually increase the amount and type of data to metadata and/or social media data in order to determine under which conditions the spoken audio can be deanonymized. A central challenge of this project is to determine how to increase the amount and type of data in an ecologically valid manner, while paying careful attention the ethical aspects of the research.

Research Question/Goal: 

Determine under which conditions deep learning techniques are able to identify speakers whose voices occur in spoken audio that has been anonymized.

Method:

The work will be carried out within the threat modelling framework for speech that was recently presented at the 4th Symposium on Security and Privacy in Speech Communication. Within this framework a series of machine-learning-based attacks will be designed and implemented, on the basis of the attacks studied in the Voice Privacy Challenge https://www.voiceprivacychallenge.org/ Careful research will be carried out in order to design an ecologically valid data set, containing speech data, but also relevant metadata. Data synthesis techniques will possibly be considered. Finally, the identification attacks will be applied to the data, which gradually increasing the amount of data assumed to be available to the attacker.

Student requirements

Experience with / interest in: Adversarial machine learning, speech processing or speech recognition, privacy, GDPR

Type of project: Master thesis

Literature

1. Cohen, A. et al., “Towards formalising the GDPRs notion of singling out,” PNAS, 2020. 
2. Rahman, M.U. et al., “Scenario of Use Scheme: Threat Model Specification for Speaker Privacy Protection in the Medical Domain”. 4th Symposium on Security and Privacy in Speech Communication, 2024. 
3. Teixeira, F. et al., "Privacy-Oriented Manipulation of Speaker Representations," in IEEE Access, vol. 12, pp. 82949-82971, 2024.
4. Tomashenko, N. et al.  The VoicePrivacy 2024 Challenge evaluation plan, 4th Symposium on Security and Privacy in Speech Communication, 2024. 
    

Supervisors  
 

Name: Marcel van Gerven 
Email: marcel.vangerven [at] donders.ru.nl 
Website: https://www.ru.nl/personen/gerven-m-van 
Department: Machine Learning and Neural Computing 
Faculty: FSW 

Name: Ahmed Elgazzar 
Email: ahmed.elgazzar [at] donders.ru.nl 
Website: https://scholar.google.de/citations?user=0lxrCcsAAAAJ&hl=en
Department: Machine Learning and Neural Computing 
Faculty: FSW 

Project Description 

Background/Motivation: 

An important goal in neuroscience is to learn models of neural information processing from data. In this project we leverage two important developments. First, the availability of unique high-dimensional multiunit data collected during naturalistic viewing tasks by our colleagues at the Netherlands Institute for Neuroscience. Second, the recent creation of a computational framework for neural data analysis developed by the Van Gerven group. With these two developments we can achieve new breakthroughs in neuroscience.  

Research Question/Goal: 

Our goal is to understand how different brain regions cooperate as naturalistic stimuli are being processed. To this end, different models of neural information processing will be developed whose parameters are estimated from data. This work contributes to the development of a new framework for neural data science as part of the Dutch DBI2 Gravitation project.  

Method: 

The computational framework embraces controlled stochastic differential equations which are estimated using variational inference. 

Student requirements 

Experience with / interest in: Students should have an interest in neuroscience and a strong background in (physics-inspired) machine learning as well as strong programming skills (Python and/or Julia). 

Type of project (master thesis, internship (only for companies)): Master thesis 

Expected time frame: February – July 2025 

Literature 

1. Ahmed Elgazzar & Marcel van Gerven. Universal Differential Equations as a Common Modeling Language for Neuroscience. https://arxiv.org/abs/2403.14510. 2024. 

2. Belinda Tzen & Maxim Raginsky. Neural Stochastic Differential Equations: Deep Latent Gaussian Models in the Diffusion Limit. https://arxiv.org/pdf/1905.09883. 2019. 

3. Le et al. MonkeySee: Space-time-resolved reconstructions of natural images from macaque multi-unit activity. https://nips.cc/virtual/2024/poster/95362. 2024.