Summer employee positions at the Department of Computer Science 2026

The Department of Computer Science is looking for summer employees

We at the Department of Computer Science want to offer motivated students a chance to work on interesting research topics with us. We are looking for BSc or MSc degree students at Aalto or other universities to work with us during the summer 2026. If you have enjoyed your studies and want to learn more about computer science, this might be your place. We do not expect you to have previous research experience; this could be the start of your bright researcher career! You will be supported by other summer employees, doctoral students and postdocs at the department.

Ready to apply?

See the complete list of the available topics below (will be fully updated by January 7th) and choose the topic(s) (max. 5) that interest you the most. Please indicate them in order of preference in the relevant section on the application form.

Please submit your application through our recruitment system. The application form will open on January 7^th and close on January 31^st, 2026, at 23:59 Finnish time (UTC +2).

Link to the application form: xxx

Are you an international student or coming from abroad?

Please check the Aalto Science Institute AScI internship programme for international summer employees.
https://www.aalto.fi/en/aalto-science-institute-asci/how-to-apply-to-the-asci-international-summer-research-programme

AScI arranges activities for international summer employees who have applied through their call and helps in finding an apartment in Espoo.

More information

If you have questions regarding applying, please contact Susanna Holma from HR Team. firstname.lastname@aalto.fi.

Summer employee topics 2026

Topics are listed here and will be fully updated by January 7^th at the latest.

1.  Adversarial Examples for Cyber-physical Vision Systems

Supervisor: Sebastian Szyller 

Contact info: seb.szyller@aalto.fi

Number of open positions: 1

The applicant is going to work with another lab member on generating adversarial examples against vision systems that aren’t purely digital — cyber-physical, e.g. adversarial examples printed out on paper, modifications to objects, or realised in AR/VR. The applicant needs to be proficient in Python including PyTorch, strong machine learning skills (specifically gradient-based optimisation) are strictly required; prior experience with computer vision is a plus.

2. EU’s Artificial Intelligence Act Role in the Regulation of Complex Systems

Supervisor: Petri Vuorimaa

Contact info: Ana Paula Gonzalez Torres, anapaula.gonzaleztorres@aalto.fi

Number of open positions: 1

The goal of the research is to understand how complex system theories can provide insights into why regulation might have disparate impacts, especially regarding impulses to “simplify regulation”. 

Preliminary research question: How can the EU’s agenda to “simplify regulation” impact the regulation of complex systems such as AI?  

Requested skills: 

-Knowledge or experience with tech regulation and policy. Keen on the EU’s AI Act. 

-Previous research experience, incl. literature reviews, research methods, academic writing. 

-General understanding of AI and current state of the art. 

-Computational social science.

3.  Computational Perspectives on Democracy

Supervisor: Corinna Coupette

Contact info: corinna.coupette@aalto.fi

Number of open positions: 1–3

Supported by the ERC Starting Grant CompLex – Toward a Computational Theory of Legal Complexity, the Telos Lab at Aalto University invites applications for several projects developing computational perspectives on democracy. While modern societies and technologies have changed dramatically over the past decades, the structure and procedures of our legal systems have remained largely unaltered. In this project, we analyze the institutions and procedures existing legal systems, investigate their shortcomings, and explore potential strategies to improve democratic processes. 

Depending on interns’ preferences and skills, topics for summer projects include: 

1. Data science for democracy and legal systems. Collect, preprocess, model, and analyze legal or political data, develop methods to study domain-specific data across multiple modalities, including networks and texts. 
   Desirable background: Experience handling complex (network and text) data, especially from law or political science.
2. Agent-based modeling of policymaking systems. Build simulations of legal systems or policymaking processes to evaluate their performance and compare them with state-of-the-art alternatives. 
   Desirable background: Experience with computational modeling and simulation design.
3. Theoretical approaches to democratic processes. Analyze the mathematical and computational properties of existing procedures for collective decision-making, e.g., regarding their power dynamics, fairness, and transparency, and investigate the properties of promising alternatives. 
   Desirable background: Experience with mathematical modeling and interest in societal applications of theoretical computer science.

For an overview of the type of work conducted in the lab, see https://www.coupette.io/publications/.
Programming languages: Python 
Libraries/frameworks for handling: dataframes, databases, network modeling and analysis, computational modeling and simulation, data analysis, data visualization, data mining, machine learning, natural language processing, web scraping. 
Potential tasks (with different emphases in different subprojects): literature review; data collection, data preprocessing, data modelling, and data engineering; computational modeling and simulation; problem formalization and theoretical analysis; method development and (experimental) validation; data analysis and visualization; software development, validation, and testing; collaboration and community engagement.

Field of study/keywords: Computational Social Science, Computational Politics, Computational Legal Studies, Legal Data Science, Computational Social Choice

4. Algorithms for information collection and explanation in logic-based automated decision-making

Supervisor: Jussi Rintanen

Contact info: jussi.rintanen@aalto.fi

Number of open positions: 1-2

Decision-making is increasingly being automated in the real-world, including in sensitive contexts like healthcare or law enforcement. It is essential that automated decisions are made in a user-friendly way. 

This includes gathering only the most relevant pieces of information related to the decisions, as well as explaining decisions to users.

In this project, we employ logic-based algorithmic methods in this domain, with the aim of transparent automated decision-making. We investigate in particular algorithms for and computational complexity of problems related to information gathering (e.g. finding the least costly order of evaluating propositions in order to arrive at a decision) and explanation (e.g. providing the least costly way to change the input features for a user that guarantees a change in the decision). These optimization problems are typically computationally hard. The project work will in large part consist of implementing efficient algorithms for these problems, with either careful custom algorithms, or algorithms using modern constraint solving, such as SAT and its optimization variants.

Necessary skills for this project include good programming skills (ideally with fast languages such as C++) and some experience with declarative problem solving. Having completed the course CS-E3220 - Declarative Programming with a good grade is considered an advantage. 

Alternatively, especially for a more theoretically motivated student, experience with computational complexity is an advantage.

5. Robustness of multimodal foundation models

Supervisor: Francesco Croce

Contact info: francesco.croce@aalto.fi

Number of open positions: 1

Despite their impressive performance across diverse tasks, modern foundation models still fail on out-of-distribution or adversarial inputs. The project focuses on analyzing failure modes of recent multimodal models, i.e. how to automatically discover and trigger such vulnerabilities, and how to design solutions. Potential topics include (the list is non-exhaustive, the specific project can be discussed): 1) adversarial robustness of methods for unlearning visual concepts in multimodal LLMs, 2) hallucinations in video LLMs, 3) interplay between compression and robustness in image tokenizers. Required: Solid foundation in machine learning fundamentals. Preferred: Hands-on experience with training/evaluating foundation models and adversarial ML techniques.

6. Bayesian workflow

Supervisor: Aki Vehtari

Contact info: aki.vehtari@aalto.fi

Number of open positions: 1

You will take part in developing computational diagnostic tools for different parts of Bayesian workflow (see, e.g. https://arxiv.org/abs/2011.01808). Possible more specific topics include model checking diagnostics, cross-validation, better priors, inference diagnostics. Prerequisites: Bayesian inference and MCMC.

7. Geometric algorithms in Euclidean and hyperbolic spaces

Supervisor: Sándor Kisfaludi-Bak

Contact info: sandor.kisfaludi-bak@aalto.fi

Number of open positions: 1

Our computational geometry research group is looking for a summer intern to help us with our research related to geometric algorithms in Euclidean and hyperbolic spaces. We expect a good understanding of mathematics (e.g. in discrete math, graph theory, geometry), algorithms, and theoretical computer science.

Potential topics include geom. optimization, parameterized geom. algorithms, fine-grained complexity, geom. approximation, spanners, geom. intersection graphs. Each of these is studied in both Euclidean and hyperbolic settings. The position is focused on fundamental theory; we prove theorems about the algorithmic properties of geometric problems.

8. Compositional Generation from Pretrained Diffusion Models

Supervisor: Polina Barabanshchikova, Samuel Kaski

Contact info: polina.barabanshchikova@aalto.fi

Number of open positions: 1

The high cost of training modern generative models, combined with the increasing availability of strong pretrained models, motivates research on model reuse and composition. Recent work has explored compositional generation using multiple pretrained diffusion models [1,2,3,4], demonstrating that combinations of specialized models allow better control over the output. We are currently completing a novel compositional framework in which different parts of a sample are generated by different diffusion models, coordinated through a game-theoretic mechanism. At present, the method is implemented only for text-to-image generation, partly because it relies on cross-attention maps derived from textual conditioning.

The goal of this internship is to generalize this compositional approach to other data modalities. As a first step, extending the method to text-to-graph [5] and/or text-to-audio [6] generation should be relatively straightforward, since these models also rely on cross-attention-based conditioning. A more ambitious objective is to remove the dependence on attention maps altogether, enabling modality-agnostic compositional generation. 

Prerequisites: Python, PyTorch, background in generative modelling

[1] Du, Yilun, et al. "Reduce, reuse, recycle: Compositional generation with energy-based diffusion models and mcmc." International conference on machine learning. PMLR, 2023.

[2] Garipov, Timur, et al. "Compositional sculpting of iterative generative processes." Advances in neural information processing systems 36 (2023): 12665-12702.

[3] Skreta, Marta, et al. "The Superposition of Diffusion Models Using the It\^ o Density Estimator." arXiv preprint arXiv:2412.17762 (2024).

[4] Bradley, Arwen, et al. "Mechanisms of projective composition of diffusion models." arXiv preprint arXiv:2502.04549 (2025).

[5] Chang, Jinho, and Jong Chul Ye. "LDMol: A Text-to-Molecule Diffusion Model with Structurally Informative Latent Space Surpasses AR Models." arXiv preprint arXiv:2405.17829 (2024).

[6] Huang, Rongjie, et al. "Make-an-audio: Text-to-audio generation with prompt-enhanced diffusion models." International Conference on Machine Learning. PMLR, 2023.

9. Meta-Learning User Models under Model Misspecification

Supervisor: Yifan Zhu, Xinyu Zhang, Samuel Kaski

Contact info: yifan.zhu@aalto.fi

Number of open positions: 1

The AI agent’s theory of mind (ToM) is an effective capability for human-AI collaboration. It enables the agent  to infer the user’s latent goals, mental states, and constraints during sequential decision-making [1]. Bayesian ToM, a common solution, inverts a structured generative model of user behavior specified a priori (e.g. computational rationality user models [2,3,4]) to infer the latent mental states and user-specific parameters from observed behavioral trajectories [5]. While such structured models provide a principled foundation for modeling human behavior, their real-world deployment often suffers from the fundamental challenge of model misspecification, where the true user’s behavioral generation process deviates from the assumed parametric model family, resulting in brittle ToM inference [6].

This internship will investigate how to combine structured user models and data-driven learning to mitigate such misspecification [7, 8]. One direction is to meta-learn a hierarchical prior and amortized posterior predictor over a family of computational rationality user models that have different cognitive structures, so that the AI agent can more robustly adapt to diverse user behaviors. The specific scope can be flexible and decided as per discussion.

Prerequisites: Python, PyTorch; background in probabilistic machine learning and deep learning; background in (deep) reinforcement learning is a plus.

[1] Çelikok, M. M., et al. (2023). Modeling needs user modeling. Frontiers in Artificial Intelligence.

[2] Zhu, Y., et al. (2025). More Than Irrational: Modeling Belief-Biased Agents. arXiv preprint arXiv:2511.12359; accepted by AAAI-26.

[3] Oulasvirta, A., et al. (2022). Computational rationality as a theory of interaction. In Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems (pp. 1-14).

[4] Howes, A., et al. (2023). Towards machines that understand people. AI Magazine, 44(3), 312-327.

[5] Baker, C., et al. (2011). Bayesian theory of mind: Modeling joint belief-desire attribution. In Proceedings of the annual meeting of the cognitive science society (Vol. 33, No. 33).

[6] Skalse, J., & Abate, A. (2023). Misspecification in inverse reinforcement learning. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 37, No. 12, pp. 15136-15143).

[7] Rabinowitz, N., et al. (2018). Machine theory of mind. In International conference on machine learning (pp. 4218-4227). PMLR.

[8] Shah, R., et al. (2019). On the feasibility of learning, rather than assuming, human biases for reward inference. In International conference on machine learning (pp. 5670-5679). PMLR.

10. Mechanistic Interpretability for Tabular Foundation Models

Supervisor: Xinyu Zhang, Conor Hassan, Julien Martinelli and Samuel Kaski

Contact info: xinyu.2.zhang@aalto.fi

Number of open positions: 1

Prior-data Fitted Networks (PFNs) are foundation models trained on synthetic datasets sampled from a prior over tasks, and then used as fast Bayesian predictors in low-data regimes [1]. For tabular problems, PFN-style transformers like TabPFN can outperform tuned classical methods on many small datasets while remaining cheap at test time [2].

Despite their success, tabular PFNs remain hard to trust and improve. Their priors are specified procedurally (e.g. structural causal models or draws from Gaussian processes over features and targets), and their internal representations over datasets and features are opaque. Two questions are central: how strongly predictions depend on the match between the pretraining prior and a given real dataset, and how PFNs internally encode task properties such as sparsity, feature relevance, correlation structure, etc.In parallel, sparse autoencoders (SAEs) have emerged as a promising tool for mechanistic interpretability in large models, discovering latent units that behave like human-meaningful “concepts’’ and can be intervened upon [3].

This internship will, to our knowledge, be the first to apply SAE-based mechanistic interpretability to tabular foundation models. We will adapt sparse autoencoder “concept discovery’’ techniques to tabular FMs, learning sparse latent codes from internal activations and probing how these codes relate to controllable generative factors in the synthetic tabular prior (e.g. sparsity, correlation structure, non-linearity, imbalance). Comparing probes on SAE codes and raw activations will indicate whether the learned units behave as disentangled “tabular concepts’’ and where in the network they are most informative, laying the groundwork for internal diagnostics and simple intervention strategies.

The discovered concepts can then be treated as intervention handles: we will perturb individual SAE units during inference to explore how they affect predictions and whether they can be used to patch specific failure modes. By varying the synthetic pretraining distribution and comparing concept-space embeddings of new datasets to those seen in pretraining, we will also explore simple internal diagnostics for prior misspecification. Because this analysis requires retraining multiple variants, we will rely on lightweight architectures, which have been shown to retain strong predictive performance [4].

The same methodology can then be applied to other foundation models, such as amortized experimental design pipelines that propose queries instead of only predicting labels [5,6,7], and tabular FMs built from dynamical systems [8,9]. With real insights into these models’ mechanisms, we could further establish a principled way of prior design to improve their robustness and generalization at inference time. 

Practical aspects and expected profile: the student will (i) review the literature on PFNs and recent work on mechanistic interpretability with SAEs, (ii) set up tabular PFN training on synthetic priors with controllable generative factors, (iii) implement SAE-based concept discovery and probing on internal activations, and (iv) use the resulting concepts to build and test diagnostics for prior mismatch and to explore simple activation-based interventions on the model. The project is ideal for a Master-level student with solid Python skills, experience with modern deep learning frameworks (PyTorch or JAX) and attention-based architectures such as Transformers, and an interest in interpretability and probabilistic modelling.

The probabilistic machine learning group at Aalto University provides an ideal environment for this project, combining expertise in Bayesian deep learning, foundation models, with close ties to real applications in biology and experimental design. In particular, the internship will be conducted in collaboration with Xinyu Zhang, Conor Hassan, and Julien Martinelli.

[1] Müller, Samuel, et al. "Position: The Future of Bayesian Prediction Is Prior-Fitted." arXiv preprint arXiv:2505.23947 (2025).

[2] Hollmann, Noah, et al. "Accurate predictions on small data with a tabular foundation model." Nature 637.8045 (2025): 319-326.

[3] Huben, Robert, et al. "Sparse autoencoders find highly interpretable features in language models." The Twelfth International Conference on Learning Representations. 2023.

[4] Pfefferle, Alexander, et al. "nanoTabPFN: A Lightweight and Educational Reimplementation of TabPFN." arXiv preprint arXiv:2511.03634 (2025).

[5] Huang, Daolang, et al. "ALINE: Joint Amortization for Bayesian Inference and Active Data Acquisition." arXiv preprint arXiv:2506.07259 (2025).

[6] Zhang, Xinyu, et al. "PABBO: Preferential Amortized Black-Box Optimization." arXiv preprint arXiv:2503.00924 (2025).

[7] Anonymous authors. “Task-Agnostic Amortized Multi-Objective Optimization”. Submitted to The Fourteenth International Conference on Learning Representations. 2025.

[8] Seifner, Patrick, et al. "Zero-shot Imputation with Foundation Inference Models for Dynamical Systems." arXiv preprint arXiv:2402.07594 (2024).

[9] d'Ascoli, Stéphane, et al. "Odeformer: Symbolic regression of dynamical systems with transformers." arXiv preprint arXiv:2310.05573 (2023).

11. Nested Theory-of-Mind Inference

Supervisor: Alex Hämäläinen, Samuel Kaski

Contact info: alex.hamalainen@aalto.fi

Number of open positions: 1

Theory of Mind (ToM) encompasses the ability of humans to attribute mental states like beliefs and goals to others and themselves. Recent advances in machine learning have made it feasible to build collaborative AI systems that are equipped with this ability; systems that can, for instance, adapt for individual users by probabilistically inferring their goals based on subtle behavioral cues. 

The fundamental engine that enables systems with such capabilities is often called a user model — a model which describes how human goal and other characteristics affect their behaviors. While recent advances in computational cognitive science have enabled specifying highly elaborate user models, capturing latent cognitive factors and biases behind human decision making, building user models with nested ToM effects has remained largely an open question. Nested ToM effects occur, for instance, when users themselves start adapting to interactive AI systems by forming nested beliefs about AI’s beliefs about themselves. The incapability to reason about such effects may result in diminished AI utility.

In this project, you will participate in our on-going cutting-edge research on building nested ToM user models and AI systems that can computationally feasibly leverage such models for better assistance. The more specific focus will be determined based on the state of the project and your interests. 

Prerequisites: Probabilistic (Bayesian) machine learning, deep learning, reinforcement learning, Python/PyTorch.

References: 

De Peuter, S., Zhu, S., Guo, Y., Howes, A., & Kaski, S. (2024). Preference learning of latent decision utilities with a human-like model of preferential choice. Advances in Neural Information Processing Systems, 37, 123608-123636.

Jha, K., Le, T. A., Jin, C., Kuo, Y. L., Tenenbaum, J. B., & Shu, T. (2024, March). Neural amortized inference for nested multi-agent reasoning. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 38, No. 1, pp. 530-537).

Oulasvirta, A., Jokinen, J. P., & Howes, A. (2022, April). Computational rationality as a theory of interaction. In Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems (pp. 1-14).

Hämäläinen, A., Çelikok, M. M., & Kaski, S. (2023, July). Differentiable user models. In Uncertainty in Artificial Intelligence (pp. 798-808). PMLR.

Çelikok, M. M., Peltola, T., Daee, P., & Kaski, S. (2019). Interactive AI with a Theory of Mind. arXiv preprint arXiv:1912.05284.

12. Bayesian physician — training a Bayesian neural network to emulate physician’s decisions

Supervisor: Jorge Loría and Samuel Kaski

Contact info: jorge.loria@aalto.fi

Number of open positions: 1

Doctors decide treatments and diagnoses in varied manners for their patients. This project aims to fit a Bayesian neural network (BNN) that can predict the treatments and diagnoses each doctor would assign to different patients. Understanding the variability of decisions between different doctors with similar patients is a task that will be studied within this project. Additionally, the obtained model will be used for a downstream task. The importance of using a Bayesian approach is that it allows to simulate from the posterior of models conditional on the observations, which is key for other downstream tasks such as fitting foundational models.

Ability to handle large volumes of data in a limited computational environment is expected.

Prerequisites: torch, variational inference, Bayesian methods, knowledge of causal inference is desirable but not expected.

References:

Zhang, C., Bütepage, J., Kjellström, H., and Mandt, S. (2018). Advances in Variational Inference. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(8):2008–2026.

Liu, Q., & Wang, D. (2016). Stein variational gradient descent: A general purpose bayesian inference algorithm. Advances in neural information processing systems, 29.

Papamarkou, T., Skoularidou, M., Palla, K., Aitchison, L., Arbel, J., Dunson, D., ... & Zhang, R. (2024). Position: Bayesian deep learning is needed in the age of large-scale AI. arXiv preprint arXiv:2402.00809.

Hollmann, N., Müller, S., Purucker, L., Krishnakumar, A., Körfer, M., Hoo, S. B., ... & Hutter, F. (2025). Accurate predictions on small data with a tabular foundation model. Nature, 637(8045), 319-326.

13. Appromximate equilibria in repeated multi-agent games

Supervisor: Alvar Haltia, Samuel Kaski

Contact info: alvar.haltia@aalto.fi

Number of open positions: 1

The combination of game theory and machine learning is a promising direction for improving real-world usage of AI. Pure game theory typically focuses on models which are too simplistic to be applicable, while machine learning has traditionally failed to properly account for strategic users and adversaries. A core problem is that computing the requisite startegies and equilibria in high-dimensional or continuous games is difficult. Standard methods often fail to capture the mixed (stochastic) strategies required for Nash Equilibria, collapsing instead to deterministic and exploitable behaviors. We are investigating methods that treat the optimal strategy not as a single point to be maximized, but as a high-dimensional distribution to be sampled. The core objective is to utilize probabilistic machine learning tools—e.g. diffusion models and flow networks—to learn policies that can represent and sample from complex, multi-modal equilibrium distributions, thereby preventing mode collapse and cycling in multi-agent interactions.

Current literature is converging on two primary directions: "Games as Inference" and "Generative Meta-Solving." The former frames finding a Nash Equilibrium as a variational inference task, where agents minimize the divergence between their policy and a target posterior representing optimality; this allows for the application of Bayesian tools to solve decentralized stochastic games. The latter replaces explicit populations of strategies (common in algorithms like PSRO) with latent generative representations. State-of-the-art methods like Generative Evolutionary Meta-Solvers (GEMS) and Diffusion Fictitious Play (DiffFP) use generative models to synthesize diverse best responses on the fly, offering a scalable alternative to maintaining large payoff matrices. We aim to apply these generative techniques to efficiently find approximate equilibria (e.g., $\epsilon$-Nash, Coarse Correlated Equilibrium) in continuous settings.

Required skills: Strong mathematical background, mainly in probability theory. Experience with probabilistic machine learning. Experience and/or confidence in implementing algorithms based on mathematical formulae. Familiarity with game theory and related concepts (Nash equilibria, etc.) useful. Familiarity with optimization algorithms in python or other common ML language.

References:

[1]A. Sharma et al., “Generative Evolutionary Meta-Solver (GEMS): Scalable Surrogate-Free Multi-Agent Learning,” Sep. 2025, doi: 10.48550/ARXIV.2509.23462.

[2]A. Karthikeyan and Y. V. Pant, “DiffFP: Learning Behaviors from Scratch via Diffusion-based Fictitious Play,” Nov. 2025, doi: 10.48550/ARXIV.2511.13186.

[3]L. M. Brunswic, H. Wang, S. Luo, J. Hao, A. Rasouli, and Y. Li, “A Theory of Multi-Agent Generative Flow Networks,” Sep. 2025, doi: 10.48550/ARXIV.2509.20408.

[4]Z. Zhao and H. Zhang, “Vairiational Stochastic Games,” Mar. 2025, doi: 10.48550/ARXIV.2503.06037.

[5]T. Freihaut, L. Viano, V. Cevher, M. Geist, and G. Ramponi, “Learning Equilibria from Data: Provably Efficient Multi-Agent Imitation Learning,” May 2025, doi: 10.48550/ARXIV.2505.17610.

14. High-dimensional Bayesian Optimization

Supervisor: Julien Martinelli and Samuel Kaski

Contact info: julien.martinelli@aalto.fi

Number of open positions: 1

High-dimensional Bayesian optimization (BO) has long been considered fragile, and much of the earlier literature proposed increasingly intricate algorithms, including random embeddings, latent subspaces, and complex surrogates. More recent work [1-3] has challenged this narrative, showing that comparatively simple BO pipelines can perform surprisingly well in high dimensions when their priors and hyperparameters are chosen carefully.

At the same time, another line of work has argued that sparsity is a central structural property of many realistic high-dimensional problems: only a small subset of coordinates, or of low-order interactions, meaningfully affect the objective. In this case, Bayesian priors that shrink irrelevant directions can dramatically improve sample efficiency [4]

This project proposes to revisit and unify these perspectives by studying how sparsity-aware structure interacts with recently proposed “simple” high-dimensional BO pipelines. We will compare surrogate models that combine additive or low-order interaction structure [5] with different sparsity-inducing priors over components. The goal is not to introduce yet another complex algorithm, but to understand when sparsity-based and additive structure actually matter for making BO work reliably in very high dimensions.

Experience with probabilistic machine learning is required. Moreover, prior familiarity with Gaussian processes, Bayesian optimization, and sparsity-inducing priors is a plus.

[1] C. Hvarfner et al., Vanilla Bayesian Optimization Performs Great in High Dimensions, 2024.

[2] L. Papenmeier et al., Understanding High-Dimensional Bayesian Optimization, 2025.

[3] C. Doumont et al., We Still Don’t Understand High-Dimensional Bayesian Optimization, 2025.

[4] D. Eriksson and M. Jankowiak, High-Dimensional Bayesian Optimization with Sparse Axis-Aligned Subspaces, 2021.

[5] X. Lu et al., Additive Gaussian Processes Revisited, 2022.

15. Probabilistic machine learning and deep generative modeling for electronic health records

Supervisor: Harri Lähdesmäki

Contact info: harri.lahdesmaki@aalto.fi

Number of open positions: 1+

We are looking for summer interns to develop novel probabilistic machine learning and generative modeling methods to analyze large-scale electronic health records datasets from biobanks and clinical trials. This project is related to our research group's general aims in developing ML and AI methods to predict adverse drug effects, survival or other outcomes using time-series health data records, which can be possibly multimodal, such as laboratory tests, diagnosis codes, text, etc. Methodologically, we are using various methods, including structured latent variable models, VAEs, Gaussian processes, transformer models, language models, and causal analysis. Studies and interest in probabilistic machine learning and generative models is expected (e.g. some of the courses: Deep Learning, Deep Generative Models, Probabilistic Machine Learning, Gaussian Processes, and/or Machine Learning in Biomedicine). Tasks for summer internship can be adapted to fit student's skills and interest. The work will be done in collaboration with other research groups from the Finnish Center for Artificial Intelligence, and the novel methods can be evaluated using real-world data sets from our collaborators in university hospitals. Work can be continued after the summer. 

More information: https://research.cs.aalto.fi/csb/

16. Deep generative modeling for dynamical systems

Supervisor: Harri Lähdesmäki

Contact info: harri.lahdesmaki@aalto.fi

Number of open positions: 1+

There is an abundance of dynamical systems around us, but many real-world systems are too complex to be modeled explicitly. Recent machine learning breakthroughs include deep learning methods for differential equations, such as Gaussian process ODEs [1], neural ODEs, neural PDEs [2], physics-guided diffusion models [7] etc. These methods are particularly useful in learning arbitrary continuous-time dynamics from data, either directly in the data space [1,3] or in a low-dimensional latent space [4,5,6]. We are looking for summer interns to join our current efforts to (i) develop efficient yet probabilistically calibrated methods to learn such black-box differential equation models directly from data (e.g. in biology, physics or video applications), and (ii) to further developing these methods for causal analysis in health and other applications. Methodologically this project can include e.g. latent variable models, neural differential equations, foundation models and amortized inference, and causal analysis. Studies and interest in probabilistic machine learning and generative models is expected (e.g. some of the courses: Deep Learning, Deep Generative Models, Probabilistic Machine Learning, and/or Machine Learning in Biomedicine). Tasks for summer internship can be adapted to fit student's skills and interest. Work can be continued after the summer.

More information: https://research.cs.aalto.fi/csb/

Our selected work:

[1] http://proceedings.mlr.press/v80/heinonen18a.html

[2] https://openreview.net/forum?id=aUX5Plaq7Oy

[3] https://arxiv.org/abs/2106.10905

[4] https://papers.nips.cc/paper/9497-ode2vae-deep-generative-second-order-odes-with-bayesian-neural-networks

[5] https://arxiv.org/abs/2210.03466

[6] https://arxiv.org/abs/2307.04110

[7] https://ml4physicalsciences.github.io/2025/files/NeurIPS_ML4PS_2025_237.pdf

17. Deep learning for single-cell biology

Supervisor: Harri Lähdesmäki

Contact info: harri.lahdesmaki@aalto.fi

Number of open positions: 1+

Single-cell sequencing technologies provide genomics data, such as gene expression and immune cell receptors, at unprecedented resolution, and deep learning methods are commonly used to analyze these datasets. We are looking for a summer intern to join our research group to develop novel probabilistic machine learning and deep learning methods for various tasks in single-cell biology, including e.g. predicting (i) patient phenotypes, (ii) treatment responses to drugs and/or genetic perturbations, and (iii) interactions between immune cell receptors and ligands. Methodologically this project can involve various neural network architechtures (attention, CNNs, GCNs, multiple instance learning), optimal transport, and generative models. Studies and interest in probabilistic machine learning and generative models is expected (e.g. some of the courses: Deep Learning, Deep Generative Models, Probabilistic Machine Learning, Machine Learning in Biomedicine, and/or Modeling Biological Networks). Tasks for summer internship can be adapted to fit student's skills and interest. Work can be continued after the summer.

More information: https://research.cs.aalto.fi/csb/

Our selected recent work:

[1] https://academic.oup.com/bioinformatics/article/41/Supplement_1/i96/8199355

[2] https://doi.org/10.1093/bioinformatics/btad743

[3] https://academic.oup.com/bioinformatics/article/39/1/btac788/6881078?login=false

[4] https://www.nature.com/articles/s41467-022-33720-z

[5] https://www.aalto.fi/en/news/an-ai-model-reveals-how-the-bodys-defence-system-recognises-skin-cancer

18. Designing for Developer Experience and Continuous Experimentation

Supervisor: Fabian Fagerholm

Contact info: fabian.fagerholm@aalto.fi

Number of open positions: 1-3

The Mind and Software research group is looking for skilled and motivated research assistants to contribute to our research on developer experience and experiment-driven software development. Developer experience refers to the cognitive, motivational, and affective experience that software developers have while developing software. We design methods and tools for studying and assessing developer experience in various contexts of modern software development. Continuous experimentation is an approach where field experiments with real users inform software development, for example, through A/B testing. We investigate methods and models for various aspects of the CE process.

We are looking for people to assist with research and software development tasks in these areas. As a research assistant, you could participate in reviewing existing scientific literature, planning and designing research studies, collecting and analysing data using qualitative and/or quantitative methods, and developing software needed for data collection in empirical studies.

Basic required skills:

- A basic understanding of empirical research on human factors in software engineering (e.g., through studies in software engineering, human-computer interaction, cognitive or social psychology, or similar).
- Ability to work in a self-directed manner as part of a team.
- Ability to read and summarise scientific literature.
- Academic writing skills (in English).

Other required skills (one or more of the following):

A. Familiarity with mobile app development using React Native, web development using React frameworks such as Next.js, and collaborative software development (ability to coordinate with team members and to use Git and continuous integration).

B. Familiarity with user interface design, creating prototypes of different fidelity (e.g., ranging from wireframes to working software prototypes), using visual design tools (e.g., Figma, Miro), and ability to create visually appealing design elements following existing design guidelines.

C. Familiarity with research methods, e.g., interviews, think-aloud, questionnaire design, specific qualitative or quantitative research methods (e.g., thematic analysis, descriptive statistics, exploratory or confirmatory data analysis methods).

In addition, we particularly appreciate an interest and motivation to work in an academic context with challenging tasks requiring flexibility in thinking combined with attention to detail.

The applicant is not required to be an expert in these areas but should display a good foundation and willingness to learn and develop their skills on their own initiative as well as in collaboration with other members of the research group.

19. Machine Learning: Foundations and New Frontiers

Supervisor: Vikas Garg

Contact info: vikas.garg@aalto.fi

Number of open positions: 2

Applications are invited for various summer employee positions in our group. An ideal student would be eager to push the frontiers of science; have strong mathematical, theoretical, statistical, or algorithmic background; and be comfortable with programming in Deep Learning. We particularly invite students with strong Math, LLM, Bioinformatics, and Physics backgrounds to apply.  See https://shorturl.at/SGagV and https://shorturl.at/4DZ4Y for experiences shared by some former members of our group. Representative publications can be found at https://people.csail.mit.edu/vgarg/

Topics of particular interest include but are not limited to:

(1) Generative Models, Neural ODEs/PDEs/SDEs, Deep Equilibrium Models, Implicit Models

(2) Large Language Models, Reinforcement Learning, Agentic AI, and AI-assisted human-guided models 

(3) Automated theorem proving and neurosymbolic reasoning

(4) Protein Design, Material Design, Molecular Simulations, and Drug Discovery 

(5) (Temporal) Graph Neural Networks, Topological Deep Learning, Topological Data Analysis (e.g., Persistent Homology) 

(6) Differential Geometry/Information Geometry/Algebraic/Spectral Methods for Deep Learning 

(7) (Approximate) Equivariant and Invariant models

(8) Fair, interpretable, and explainable methods

20. Diplomityöntekijä, digitaalisten terveyspalveluiden käytettävyys ja saavutettavuus

Supervisor: Sari Kujala

Contact info: sari.kujala@aalto.fi

Number of open positions: 1

Diplomityön aihe liittyy digitaalisten terveyspalveluiden käytettävyyteen ja saavutettavuuteen. Vaatimuksena on HCI-alan menetelmien hallitseminen ja suomen kielen osaaminen. Pääaineena mielellään HCI tai Informaatioverkostot ja tutkimusmenetelmäkurssin suoritus katsotaan eduksi.

21. Machine-learning aided simulations of magnetised fluids mimicking the Sun

Supervisor: Maarit Korpi-Lagg

Contact info: maarit.korpi-lagg@aalto.fi

Number of open positions: 1

Solar eruptions, known as bad space weather, have a direct origin in the solar internal magnetism. How exactly solar eruptions are driven from the underlying internal magnetism is poorly known. In this project you will contribute to significant improvements of modelling the roots of bad space weather. 

The mainstream direct numerical simulations (DNS) of the strongly turbulent and magnetised solar interior flows are prohibitively expensive due to the need to resolve the smallest spatial scales. An alternative approach, Large Eddy Simulations (LES), gets around this by not resolving the smallest scales but by spatially averaging them and explicitly modeling the effect of the unresolved small-scale turbulence (SGS models). A modern approach for SGS modeling is to use machine learning (ML) to learn the effect of the unresolved turbulence from the existing averaged flow. 

So far, we have developed two different kinds of ML approaches to measure the SGS effects: CNNSs and Fourier Neural Operators. While the results are, in general, rather satisfactory, we would like to extend the work towards a more general framework, where also the physical constraints would be taken into account. This means that the existing methods would need to be developed towards physics-informed neural networks (PINNs), in which also some well-defined symmetry conditions would be satisfied. This would be the task of the summer intern.

22. Distributed quantum computing

Supervisor: Jukka Suomela

Contact info: jukka.suomela@aalto.fi

Number of open positions: 1

Our research group "Distributed Algorithms" is looking for a summer intern to help us with our research related to distributed quantum algorithms. Our work is at the intersection of mathematics (especially discrete mathematics and graph theory), theoretical computer science (especially theory of distributed computing and distributed graph algorithms), and quantum physics (especially quantum information theory). For more information on our research and our research group, see https://research.cs.aalto.fi/da/

23. Software developers for our course infrastructure

Supervisor: Jukka Suomela

Contact info: jukka.suomela@aalto.fi

Number of open positions: 1-3

We are hiring summer interns to help us with the development of the computer systems that keep our courses "Computer as a Tool", "Programming Parallel Computers", "Competitive Programming", and "Distributed Algorithms" up and running. We are looking for summer interns who have got strong programming skills. Some prior experience with developing course automation systems is a plus. For more information on these courses, see: https://lapio.cs.aalto.fi https://ppc.cs.aalto.fi https://plus.cs.aalto.fi/cs-e4595/ https://jukkasuomela.fi/da2020/

24. Integrating a fast ad-hoc RDB manager to the R statistics and visualisation package

Supervisor: Pekka Orponen

Contact info: pekka.orponen@aalto.fi

Number of open positions: 1

We have recently developed a fast cross-tabulation based technique for performing SQL queries on an ad-hoc RDB built from a collection of CSV data files. The technique supports also non-standard SQL queries that contain set operations (ANY, ALL, EXACTLY, ONLY, NONE). The current software runs on the user's local workstation/laptop on top of the lightweight, serverless DuckDB database management system. The proposed summer project concerns interfacing this database management and query tool to the R statistical computing package for further analysis and visualisation of the data. The key parts of the project are: (1) providing a simple tool for the user to create the cross-tabulated RDB files from their CSV data; (2) setting up the query interface using the R client for DuckDB; (3) developing a preprocessor ("transpiler") for converting complex queries containing set operations to standard SQL.

The work can be pursued in either a Linux, MacOS or Windows environment, contingent on the availability of the required software components.

The project requires familiarity with relational databases and the SQL query language, together with good programming skills preferably in Python or Julia. For task (3) also some knowledge of basic compiler techniques is needed, e.g. from course CS-C2160/CS-C2161. Previous experience in using the R package is an additional asset.

25. Boolean network-informed Gaussian processes for biological dynamical systems

Supervisor: Julien Martinelli, Harri Lähdesmäki

Contact info: julien.martinelli@aalto.fi

Number of open positions: 1

We are looking for a summer intern to develop probabilistic machine learning methods that combine Boolean regulatory networks with data-driven models of biological dynamics. Many biological systems are only partially understood: we may know which genes or proteins regulate each other, and whether they act through “AND/OR/NOT” logic, but not the detailed kinetic equations. In this project, we will explore how such qualitative regulatory knowledge can be translated into structured priors or architectures for Gaussian process–based dynamical models and, optionally, Neural ODEs. The work will include reviewing the literature on data-driven models for dynamical systems, implementing baseline GP-ODE or Neural ODE models for small regulatory networks, designing Boolean-informed variants, and comparing how they perform in terms of data efficiency, extrapolation, and uncertainty calibration under noisy, partially observed conditions. The project is ideal for a Master’s student with a strong background in (probabilistic) machine learning, solid Python programming skills (e.g. PyTorch, JAX, GPyTorch), and an interest in dynamical systems and computational biology. No prior biology knowledge is required; curiosity about biological applications is a plus. Work can be continued after the summer.

26. Massively Parallel Algorithms for Graph Problems

Supervisor: Jara Uitto

Contact info: jara.uitto@aalto.fi

Number of open positions: 2

Parallel processing of data and distributed computing are gaining attention and becoming more and more vital as the data sets and networks we want to process are overgrowing the capacity of single processors. To understand the potential of modern parallel computing platforms, many mathematical models have emerged to study the theoretical foundations of parallel and distributed computing. In this project, we study algorithm design in these models with a particular focus on the Massively Parallel Computing (MPC) and Local Computation Algorithms (LCA) models.

The problems we study are often in (but not limited to) the domain of graphs, that serve as a very flexible representation of data. We are interested in, for example, the computational complexities of classic problems such as finding large independent sets, matchings, flows, clustering problems, etc.

The applicant is assumed to have a solid knowledge of mathematics, knowledge on the basics of graph theory, and a good command of English. No prior knowledge in distributed computing is required, although it might be helpful.

27. Internship positions in probabilistic machine learning and generative AI

Supervisor: Arno Solin

Contact info: arno.solin@aalto.fi

Number of open positions: 2

We are looking for motivated interns to join ongoing research projects in probabilistic machine learning, with opportunities across: (1) Structured and efficient inference and generative modelling (mostly diffusion models), (2) uncertainty quantification (UQ) in deep learning (including large-scale models), and (3) multi-modal modelling (mostly VLMs), including semantic understanding + 3D scene reconstruction (Gaussian splatting).

You will work on cutting-edge research problems such as uncertainty-aware neural networks, approximate inference methods, and probabilistic approaches to vision–language and 3D reconstruction. The projects are part of a broader initiative to advance the foundations and practice of modern machine learning.

A strong candidate typically has: (1) Background in probabilistic modelling and approximate inference; (2) solid understanding of machine learning and deep learning; (3) strong programming skills in Python and experience with at least one of PyTorch/JAX; and most importantly (4) curiosity, creativity, and a collaborative mindset.

In your application, briefly highlight: (1) Your relevant coursework and/or research experience; (2) the topics you are most excited about (UQ, vision–language, 3D, etc.); and (3) links to code, papers, or projects (if available)

For representative publications and background, see the supervisor’s homepage: https://arno.solin.fi

28. Social media data analysis

Supervisor: Mikko Kivelä

Contact info: mikko.kivela@aalto.fi

Number of open positions: 1

We are looking for interns to join our research group who is using a computational social science approach to analyse social media. Our work spans from characterising and quantifying polarisation to detecting coordinated behavior and information operations. The methodology we employ is large based on network analysis and natural language processing. We have previously worked on Twitter/X and Reddit and are currently mostly working on Bluesky, but we are also interested in other platforms. The topic can be tailored to the interests and skill set of the intern. We will provide large-scale data sets and various starting points for your research based on them, and in the ideal case the intern will build on them based on their own ideas. The intern will be primarily adviced by a senior PhD student or a postdoc, but is encouraged to interact and collaborate and seek support with the rest of the research group. We encourage you to look at the recent articles published by our research group for examples of the type of research we are doing.

At minimum the intern should have excellent programming skills and some experience in data science. An ideal candidate will also have more experience at least on some of the methodologies we are using, such as network analysis, NLP, or social media analysis.