Home » Admissions

Admissions

We offer the possibility to join world-class research groups either at Aalto University or University of Helsinki, with multiple interesting research projects to choose from. We organise two open calls for new doctoral students annually.

Our fully-funded doctoral student positions have gained a lot of international interest since their introduction in 2014. We appreciate applicants with diverse backgrounds and enthusiasm in computer science.

The quality of research and education in both HICT universities is excellent and our doctoral students are typically hired as full-time employees for the duration of their doctoral studies.

You are most welcome to apply for HICT doctoral student positions! Our open positions are announced on this webpage.

Summer call 2021 – How to apply?

  1. Get familiar with the application process described on General HICT Call FAQ.
  2. Choose the projects from the list below.
  3. Fill in the application form and send it before deadline.

The next HICT call will be open in July 2021.

The application form closes on August 16, 2021 at midnight Finnish time (23:59 EET, Eastern European Time).

Projects in the previous HICT Spring call 2021 (call closed)

FCAI topics

Finnish Center for Artificial Intelligence FCAI is a community of experts that brings together top talents in academia, industry and public sector to solve real-life problems using both existing and novel AI. FCAI’s research mission is to create a new type of AI that is data efficient, trustworthy, and understandable. We aim to build AI systems capable of helping their users in AI-assisted decision-making, design and modeling. We develop AI techniques needed for systems which can help their users make better decisions and design better solutions across a range of tasks from personalized medicine to materials design. A core insight in developing such AIs is that they need to have world models for understanding the world and interacting with it, and user models for understanding the user and interacting with them. Read more about our research here.

Choose your topics below and fill in the application form.

Project 1. Large-scale computing on modern architectures and systems: Programming models, Designs and Optimization

Supervisor: Prof. Hong-Linh Truong, Aalto Systems and Services Engineering Analytics (http://rdsea.github.io) (Department of Computer Science, Aalto University)

The topic will focus on researching and developing techniques and tools for large-scale computing on modern architectures and systems. We will investigate emerging programming models, such as function-as-a-service, hybrid workflows of traditional data analysis pipelines and machine learning pipelines, and human-in-the-loop in large-scale, complex analysis. Novel software techniques and methods will be developed for support the developer to design, manage and optimize large-scale computing and data analysis applications. Observability, performance, elasticity, fairness and interpretability will be particular features in the focus of the design and optimization.

Large-scale computing will be targeted to emerging solutions for large-scale and data-intensive applications, which combine traditional workflow-based large-scale data analysis with machine learning, whereas modern systems will be high-performance computing systems with CPU/GPU, large-scale containerized systems, systems with AI accelerators, and potentially quantum computing systems.


Project 2. Doctoral students for ML Engineering research

Supervisors: Prof. Jukka K Nurminen (jukka.k.nurminen@helsinki.fi) Prof. Tommi Mikkonen (tommi.mikkonen@helsinki.fi), Department of Computer Science, University of Helsinki

We are looking for doctoral students to work on tools and methodologies for the software engineering of machine learning systems. To ensure that machine learning systems work for real, new ways are needed to ensure their correct and efficient operation as well as their smooth development, monitoring, and maintenance. At the moment we are running or about to start multiple European and national projects focusing on testing of AI systems, MLOps, and big data analytics.

The aim of our empirical and experimental approach is to come up with new and improved solutions for ML system development and operation. The candidate is expected to analyze, measure, and model alternative approaches and create new ideas and insights based on those. This includes implementing research prototypes to try out ideas and to collect and analyze data. Documenting the results in scientific papers is naturally important. In addition to research, the doctoral student is expected to contribute to other common academic tasks such as teaching, generation of new research ideas, and writing grant proposals.


Project 3. Query processing and optimization for next-generation databases

Supervisor: Prof. Jiaheng Lu (Department of Computer Science, University of Helsinki)

As more businesses realized that data is critical to making the best possible decisions, we see the continued growth of systems that support a massive volume of non-relational or unstructured data. We are looking for a new Ph.D. student to develop novel algorithms and theories for a unified database management system to manage both well-structured data and NoSQL data in the era of big data. Link to a webpage with further information: https://www.helsinki.fi/en/researchgroups/unified-database-management-systems-udbms


Project 4: Atmospheric AI (The exploratory data analysis group)

Supervisor: Assoc. Prof. Kai Puolamäki (Department of Computer Science and Institute for Atmospheric and Earth System Research (INAR), University of Helsinki)

The exploratory data analysis group is looking for a doctoral student for a project that focuses on the use of artificial intelligence and machine learning on natural sciences, especially in atmospheric and earth sciences. The objective is to build probabilistic models of simulated and measured phenomena and develop methods of interactive artificial intelligence that would enable the substance area experts to build these models. The project can be tailored to focus more on computer science or atmospheric sciences, depending on qualifications and preferences of the applicant. The project will be done in collaboration with the Institute for Atmospheric and Earth System Research (INAR) at the University of Helsinki. Please contact Prof. Kai Puolamäki at kai.puolamaki@helsinki.fi for further information.


Project 5: Theory of safe algorithms and their applications in Bioinformatics

Supervisor: Assoc. Prof. Alexandru Tomescu (Department of Computer Science, University of Helsinki) https://www.cs.helsinki.fi/u/tomescu/

For an algorithmic problem admitting multiple solutions, a “safe algorithm” is one reporting only those partial solutions that are common to *all* solution to the problem. Thus, the output of a safe algorithm is most likely correct. This is a novel algorithmic perspective from which to tackle both theoretical and practical problems, and could revolutionize the field of Bioinformatics.

We are looking for a PhD student to study safe algorithms for both (1) string problems (aligning biological sequences, with applications in understanding protein function across the tree of life), and (2) graph-theoretic problems, with applications in assembling viral strains (such as for SARS-CoV-2 data). We expect the candidate to have expertise in both algorithms & theory and programming, through previous relevant coursework or projects.

The PhD student will join the Graph Algorithms team (https://www.helsinki.fi/en/researchgroups/algorithmic-bioinformatics/teams/graph-algorithms), which is part of the wider Algorithmic Bioinformatics group at the University of Helsinki.


Project 6: Probabilistic real-time machine learning

Supervisor: Prof. Arno Solin (Department of Computer Science, Aalto University)

We are looking for exceptional and highly motivated doctoral students to work on algorithms and applications for real-time machine learning. Central topics and themes in this project include approximate inference methods, stochastic differential equations (incl. neural SDEs), state space modelling, and Gaussian processes. Applications of interests are in online decision-making, sensor fusion, audio/video analysis, control (also linking to RL), and robotics. An intuition of the theoretical backdrop is provided in this ICML tutorial (https://youtu.be/vTRD03_yReI) and this ICML paper (https://arxiv.org/abs/2007.05994).

The work will be done in close collaboration with the supervisor and other members of the team at Aalto University. Doctoral students in the group are encouraged to make research or internship visits to collaborating universities/companies during the course of study. Successful candidates are expected to have completed a Masters degree and have familiarity with machine learning and statistics.

For more information and recent publications and pre-prints, see the research group home page at http://arno.solin.fi


Project 7: Deep learning with probabilistic principles

Supervisor: Prof. Arno Solin (Department of Computer Science, Aalto University)

We are looking for exceptional and highly motivated doctoral students to work on algorithms and methods in combining probabilistic methods with deep learning. This project relates to various topics in this space: Bayesian deep learning, computer vision, generative models, meta-learning, Gaussian and neural processes, and normalizing flows. This project is part of a consortium project with Tampere University, with collaborators in Cambridge, Oxford, Prague, and Moscow. Examples of recent work in the group in this area include this ICCV paper (https://aaltoml.github.io/GP-MVS), this NeurIPS paper (https://arxiv.org/abs/1912.10321), and this NeurIPS paper (https://arxiv.org/abs/2010.09494).

The work will be done in close collaboration with the supervisor and other members of the team at Aalto University. Doctoral students in the group are encouraged to make research or internship visits to collaborating universities/companies during the course of study. Successful candidates are expected to have completed a Masters degree and have familiarity with machine learning and statistics.

For more information and recent publications and pre-prints, see the research group home page at http://arno.solin.fi


Project 8. Deep learning with differential equations

Supervisor: Markus Heinonen (Department of Computer Science, Aalto University)

We are looking for an exceptional and motivated phd student to push the boundaries of deep learning with differential equations. In conventional deep learning the inputs are transformed by a sequence of layers, while an alternative paradigm emerged recently interpreting learning tasks as continuous flows with ODEs or SDEs. We aim at developing new ways to perform machine learning by repurposing differential equations. Topics range from developing interpretable neural ODEs for supervised tasks, to modelling distributional structures and data augmentations with SDEs and PDEs, to adversarial, robust or probabilistic ODEs.

The work continues on the foundations of our NIPS’19 and AISTATS’19 publications “ODE2VAE” and “Differential Gaussian processes“. The work will be done in collaboration with several researchers studying the topic at Aalto and internationally. Sufficient familiarity with statistics, math, physics or machine learning is advantageous.


Project 9: Sample-efficient probabilistic machine learning

Supervisor: Prof. Luigi Acerbi (Department of Computer Science, University of Helsinki)

We are looking for an exceptional doctoral student eager to work on new machine learning methods for smart, robust, sample-efficient probabilistic inference, with applications in scientific modeling (e.g., computational and cognitive neuroscience). In our group, we are interested in developing novel approaches for building approximate Bayesian posteriors using only a small number of likelihood evaluations, which can be a game-changer for complex models or when resources are limited. Think of Bayesian optimization, but scaled up to full Bayesian inference. A state-of-the-art framework being developed in our group is Variational Bayesian Monte Carlo (VBMC), which combines Gaussian process surrogates, active learning, variational inference and Bayesian quadrature (Acerbi, NeurIPS, 2018, 2020). 

Promising thesis projects include extending the representational power of VBMC (e.g., discrete variables, more complex posteriors, higher dimension); exploiting recent advances in Gaussian process inference for superior scalability; and exploring the theoretical properties of the framework. Successful candidates are expected to have completed a Masters degree and have familiarity with machine learning and Bayesian statistics.

For more information and recent publications, see the research group home page at http://www.helsinki.fi/machine-and-human-intelligence


Project 10. Probabilistic modelling and Bayesian machine learning

Supervisor: Prof. Samuel Kaski, (Department of Computer Science, Aalto University)

I am looking for a doctoral student eager to join the Aalto Probabilistic Machine Learning Group, to work on new probabilistic models and inference techniques. Particularly promising thesis topics are: (1) simulator-based inference, for combining first-principles models with learning from data, (2) Bayesian deep learning, (3) Bayesian reinforcement learning and inverse reinforcement learning, (4) multi-agent modelling, and (4) privacy-preserving machine learning and synthetic data generation. The group has excellent opportunities for collaboration with topnotch partners in multiple applications, including user interaction, health, genomics, and neuroscience.

Links: http://research.cs.aalto.fi/pml


Project 11: Improve drug design with human assisted AI

Supervisors: Prof. Samuel Kaski, Research Fellow Markus Heinonen (Department of Computer Science, Aalto University)

Not all relevant knowledge is explicitly accessible and usable for machine learning modelling in drug design. A new and emerging area in machine learning is knowledge elicitation from human experts to improve the prediction accuracy of models so called human-in-the-loop-modelling. The goal of this PhD studentship is to develop human-in-loop machine learning models applicable for drug design. 

The PhD student will join an ongoing collaboration between Aalto and AstraZeneca (Gothenburg, Sweden). The main task will be to develop human-in-the-loop modelling so it can be used to guide deep learning based drug design. The student will be developing systems to query drug experts for information to improve machine learning models. The candidates are expected to have background in machine learning, statistics or mathematics, with keen interest on learning chemistry.


Project 12: Decomposable latent representations for toxicity prediction

Supervisors: Prof. Samuel Kaski, Research Fellow Markus Heinonen (Department of Computer Science, Aalto University)

We are looking for an exceptional and motivated phd student to develop machine learning methods for drug discovery, with a specific target of developing decomposable latent representations for in-vivo toxicity prediction. We aim at ameliorating the key challenges of drug modelling under scarce data by developing interpretable generative models with intuitive chemical or clinical explanations. The developed methods will be based on Bayesian and deep learning principles, following the successes of VAEs and GANs.

The research will be conducted and supervised jointly by the two sites of Aalto and Janssen pharmaceuticals (Belgium) under an ongoing collaboration. The candidates are expected to have background in machine learning, statistics or mathematics, with keen interest on learning chemistry.


Project 13: Bayesian deep learning

Supervisors: Prof. Samuel Kaski (samuel.kaski@aalto.fi), Research Fellow Markus Heinonen (markus.o.heinonen@aalto.fi) (Department of Computer Science, Aalto University)

We are looking for a doctoral student to join the Aalto Probabilistic Machine Learning Group, to work on developing state-of-the-art Bayesian deep learning. Key research questions are about more useful neural parameterisations, process priors on function spaces and more efficient probabilistic inference methods for deep neural networks. Possible applications range from large-scale image classification to sample-efficient Bayesian reinforcement learning and robotics.

This work will build on top of existing research lines in the group on RL and BNNs, with a recent highlight work of implicit BNNs with state-of-the-art ImageNet performance while maintaining Bayesian principles. The group has excellent collaboration and application opportunities. Background in machine learning, statistics or math is expected.


Project 14: Mobile Cross Reality through Immersive Computing (MeXICO)

Supervisors: Prof. Mario Di Francesco, Matti Siekkinen (Department of Computer Science, Aalto University)

The MeXICO project aims to overcome the limited resources of mobile devices for novel applications in cross-reality (XR): virtual, augmented, and mixed reality. Current solutions for interactive and mobile XR require real-time rendering but they are constrained by the limited resources of mobile devices. A promising approach to overcome this issue is to offload most of the heavy computing tasks from the mobile device to a remote processing unit in the cloud or at the edge.

Unfortunately, this approach faces challenges in terms of both latency and bandwidth. In fact, a noticeable motion-to-photon latency is highly detrimental in interactive applications, as it may even cause severe discomfort in addition to a poor user experience. Moreover, transmitting high-quality graphic content from remote servers to mobile devices requires a large amount of network bandwidth. Motivated by these challenges, MeXICO explores solutions for mobile and distributed XR to enable novel and effective applications.


Project 15: HAIC: Open doctoral student position in Prof. Janne Lindqvist’s group – security engineering and usable security

Supervisor: Prof. Janne Lindqvist https://www.aalto.fi/en/janne-lindqvist (Aalto University, Department of Computer Science, Director Helsinki-Aalto Institute for Cybersecurity (HAIC) HAIC.fi)

We are looking for PhD students interested in security engineering, usable security and human-computer interaction. Background and interest in systems security, security engineering, data science, machine learning, modeling, human-computer interaction or social and behavioral sciences is required. PhD topic will be agreed together with the applicant. Examples of work done in the group can be found at the website for the group https://www.lindqvistlab.org/. The students will also get to participate in the activities of HAIC. Please contact me at the aalto.fi email address about these positions.


Project 16: Open doctoral student position in Prof. Janne Lindqvist’s group – understanding video streaming user experiences

Supervisor: Prof. Janne Lindqvist https://www.aalto.fi/en/janne-lindqvist (Aalto University, Department of Computer Science, Director Helsinki-Aalto Institute for Cybersecurity (HAIC) HAIC.fi)

We are looking for doctoral students interested in understanding video streaming user experiences. Background and interest in measuring user experience, modeling, human-computer interaction, computer science or social and behavioral sciences is required.  Examples of work done in the group can be found at the website for the group https://www.lindqvistlab.org/. Please contact me at the aalto.fi email address about these positions.


Project 17: HAIC: Open doctoral student position in Prof. Janne Lindqvist’s group – artificial intelligence and machine learning for systems security and privacy

Supervisor: Prof. Janne Lindqvist https://www.aalto.fi/en/janne-lindqvist (Aalto University, Department of Computer Science, Director Helsinki-Aalto Institute for Cybersecurity (HAIC) HAIC.fi)

We are looking for doctoral students interested in developing novel artificial intelligence and machine learning approaches to security engineering and systems security and privacy. Background and interest in data science, machine learning, statistics and computational approaches to computer science are required. Examples of work done in the group can be found at the old website for the group https://www.lindqvistlab.org/. The students will also get to participate in the activities of HAIC.

Please see specific examples also http://jannelindqvist.com/publications/IMWUT19-fails.pdf http://jannelindqvist.com/publications/NDSS19-robustmetrics.pdf Please contact me at the aalto.fi email address about these positions.


Project 18: Open doctoral student position in Prof. Janne Lindqvist’s group – multitasking and productivity tools

Supervisor: Prof. Janne Lindqvist https://www.aalto.fi/en/janne-lindqvist (Aalto University, Department of Computer Science, Director Helsinki-Aalto Institute for Cybersecurity (HAIC) HAIC.fi)

We are looking for doctoral students interested in understanding productivity tools and multitasking. Background and interest in measuring user experience, modeling, human-computer interaction, computer science or social and behavioral sciences is required. Examples of work done in the group can be found at the website for the group https://www.lindqvistlab.org/. Please contact me at the aalto.fi email address about these positions.


Project 19: Open doctoral student position in Prof. Janne Lindqvist’s group – mixed methods HCI and security research

Supervisor: Prof. Janne Lindqvist https://www.aalto.fi/en/janne-lindqvist (Aalto University, Department of Computer Science, Director Helsinki-Aalto Institute for Cybersecurity (HAIC) HAIC.fi)

We are looking for doctoral students interested in pushing the envelope in mixed methods HCI and security research. Background and interest in measuring either qualitative methods or quantitative methods, and interested to learning new methods, user experience, modeling, human-computer interaction, computer science or social and behavioral sciences is required.  Examples of work done in the group can be found at the website for the group https://www.lindqvistlab.org/. Please contact me at the aalto.fi email address about these positions.


Project 20: Deep learning for electronic health records

Supervisor: Prof. Pekka Marttinen (pekka.marttinen@aalto.fi) (Department of Computer Science, Aalto University)

We will develop novel deep learning models for healthcare time series data. This work is done in collaboration with the Finnish Institute for Health and Welfare and other healthcare data holder in the Helsinki region. In particular, we will focus on 1) probabilistic deep learning, to address uncertainty in the predictions, 2) causal inference, required for decision making on individual and population levels, and 3) interpretability, which helps communicate the results to patients and policy makers. Also important are topics related to privacy and anonymization. The advances made in the project are central for the trustworthiness and acceptability of the methods in practice.


Project 21. Theoretical Framework and Deep Learning Algorithms for Large Output Spaces

Supervisor: Prof. Rohit Babbar (Department of Computer Science, Aalto University)

There is a growing interest in supervised learning for classification of data with large number of outputs or labels. However, there still lacks good theoretical understanding of algorithms in this domain with only few works such as [1,2]. Unlike other domains, the algorithmic success of deep learning methods for large output spaces has also been somewhat limited. The goal of the PhD thesis would be to explore novel theoretical frameworks and deep architectures in this domain. The scope of the project also includes exploring connections with adversarial robustness analysis of learning algorithms [3,4]. Requirements: Background in Linear Algebra, probability and Optimization. Programming experience with python. Deep learning frameworks such as PyTorch, and working with large datasets. 

  • [1] Stochastic Negative Mining for Learning with Large Output Spaces, AISTATS 2019
  • [2] A no-regret generalization of hierarchical softmax to extreme multi-label classification, NIPS 2018
  • [3] Robustness May Be at Odds with Accuracy, ICLR 2019
  • [4] Data scarcity, robustness and extreme multi-label classification, Machine Learning Journal, 2019

Project 22. Reconstructing Crisis Narratives for Trustworthy Communication and Cooperative Agency

Supervisor: Nitin Sawhney, Professor of Practice, Department of Computer Science, Aalto University

The research project, jointly conducted between Aalto University and the Finnish Institute for Health and Welfare (THL), proposes to analyze and reconstruct crisis narratives using mixed-methods, combining qualitative research for narrative inquiry with computational data analytics of crisis discourses in news and social media to understand global pandemics. Candidates will work at the intersection of Human-Computer Interaction (HCI), design research, computational social sciences, and public health for critical societal impact. We expect the candidate to have background in computer science, media and communication studies, social science, or similar disciplines.

Potential duties and tasks may include the following (to be conducted as part of the research team):

1. Examining the narratives emerging in crisis-related communication using qualitative research methods across various data sources including organization communications, news/media coverage, and social media exchange among diverse publics.

2. Automating content analysis for narrative work using suitable machine learning techniques for Natural Language Processing (NLP) such as Conversation Analysis, Content Classification, and/or Sentiment Analysis. This includes collecting and curating datasets, devising suitable methodologies, setting up the research infrastructure and tools, and a pipeline for data extraction, analysis and validation.

3. Representing and visualizing crisis narratives to support understanding and collaborative sensemaking among key stakeholders and diverse publics. These tools should support browsing, searching, and exploring information for selected crisis themes and narratives. Work in this area includes not only developing prototypes of visualizations, but also conducting design research, user experience (UX) evaluation, and pilot assessment of such tools.

The candidate is not expected to master all these domains, but work closely with a multi-disciplinary research team to lead design and development efforts, while learning and contributing to ongoing work in specific research areas of interest.

The position belongs to the Aalto career system and the selected person will be appointed for a two-year fixed term appointment with an option for renewal.


Project 23. Open doctoral student position in Keijo Heljanko’s group – Massively parallel distributing computing

Supervisor: Prof. Keijo Heljanko, Department of Computer Science, University of Helsinki and HiDATA Helsinki Centre for Data Science (https://www.helsinki.fi/en/helsinki-centre-for-data-science)

The PhD Student position is in the context of the Academy of Finland project “Design and Verification Methods for Massively Parallel Distributed Systems (DeVeMaPa)”. The main focus of the project is on the use of GPU computing to accelerate Big Data processing and the theoretical foundations of these systems.

The project will develop methodology for the design and verification of massively parallel heterogeneous computing. We need new methods to support the massive increase in the amount of parallelism at all levels of the hardware/software stack. Such massive increases in parallelism will make some currently used programming paradigms infeasible and thus new methods need to be devised to cope with industrial Big Data use cases. These methods must also be accompanied with solid theoretical foundations, allowing for the development of automated testing and verification tools that are required to validate the parallelization runtime software before production deployment. A key challenge is the need for seamless integration of heterogeneous computing with GPUs and hardware accelerators (e.g., neural network accelerators), how can they all be handled in a unified Big Data programming framework?


FCAI topics

Finnish Center for Artificial Intelligence FCAI is a community of experts that brings together top talents in academia, industry and public sector to solve real-life problems using both existing and novel AI. FCAI’s research mission is to create a new type of AI that is data efficient, trustworthy, and understandable. We aim to build AI systems capable of helping their users in AI-assisted decision-making, design and modeling. We develop AI techniques needed for systems which can help their users make better decisions and design better solutions across a range of tasks from personalized medicine to materials design. A core insight in developing such AIs is that they need to have world models for understanding the world and interacting with it, and user models for understanding the user and interacting with them. Read more about our research here.

FCAI topics within the present HICT call are listed below (projects 24-40).


Project 24: Computational modeling of human motivation and experience

Supervision: Antti Oulasvirta (Aalto University); potential other supervisors: Perttu Hämäläinen (Aalto University), Arto Klami (University of Helsinki), Elisa Mekler (Aalto University)

One of our aims at FCAI is ‘artificial human understanding’, a type of AI that much better estimate what human collaborators want or prefer or experience. The goal is to estimate the deeper (latent) motivational processes and subjective constructs like perceived aesthetics, and thereby offer help to human collaborators more effectively. We are looking for an outstanding researcher to help us construct computational models of motivations and experience rooted in most current psychological theory, and use those models to boost the inferential capability of AI assistants. In particular, we will model designers as agents that maximize expected utility in their choices but are limited by their own bounds and that of the design tools (see Gershman et al. Science 2015). This area, called computational rationality, is an exciting convergence point of machine learning and cognitive science. The ideal candidate has demonstrated track record modeling user behavior or cognition, with publications on reinforcement learning, probabilistic models, or cognitive models.

Keywords: computational modeling, human motivation, computational rationality, interactive AI, reinforcement learning, cognitive models

Previous papers: Roohi, Shaghayegh, et al. “Predicting Game Difficulty and Churn Without Players.” Proceedings of the Annual Symposium on Computer-Human Interaction in Play. 2020.


Project 25: Computational cognitive models for AI assistance

Supervision: Antti Oulasvirta (Aalto University); Samuel Kaski (Aalto University)

A prime goal for FCAI’s research is to develop a new form of AI that can better work with people and assist them in everyday tasks. We believe that deep integration of human cognition into the technical principles that govern the AI’s operation is necessary for ethically acceptable applications. We are now looking for an outstanding, methodologically oriented candidate to help us develop computational cognitive models based on the theory of computational rationality (Gershman 2015 Science). The ideal candidate has previous background in MDP/POMDP-based modeling, decision theory, multi-agent systems, or reinforcement learning.

Previous papers:

  • Kangasrääsiö, Antti, et al. “Parameter inference for computational cognitive models with Approximate Bayesian Computation.” Cognitive Science 43.6 (2019): e12738.
  • Gebhardt, C., Oulasvirta, A., & Hilliges, O. (2020). Hierarchical Reinforcement Learning as a Model of Human Task Interleaving. arXiv preprint arXiv:2001.02122.

Keywords: Computational cognitive modeling, AI assistance, interactive AI


Project 26: AI-assisted modelling of dynamic interactional data

Supervision: Samuel Kaski (Aalto University), Vikas Garg (Aalto University) and Arno Solin (Aalto University)

This project is concerned with modelling on graphs where the environment is uncertain and can change, possibly due to user interaction [1]. Graph neural networks, the state-of-the-art models for embedding graphs, have almost exclusively focused on non-strategic settings. However, several important applications involve agents on networks that compete/collaborate [2] as part of decision-making. This project is concerned with modeling the uncertainties due to the different players. We are looking for doctoral students to work on the intersection of graph neural networks, Gaussian processes, and dynamical systems (see [3] for an overview). The ideal candidate will have a strong mathematical/statistical training and good programming skills.

References:

  • [1] Ruotsalo, Tuukka, Jacucci, Giulio, Myllymäki, Petri, and Kaski, Samuel. ”Interactive intent modeling: information discovery beyond search”. Communications of the ACM, 58(1):86-92, 2015.
  • [2] Garg, Vikas, and Jaakkola, Tommi. “Predicting deliberative outcomes.” In International Conference on Machine Learning, pp. 3408-3418. PMLR, 2020. Link: https://www.mit.edu/~vgarg/DeliberativeOutcomes_CameraReady.pdf
  • [3] Solin, Arno. “Machine Learning with Signal Processing”. ICML 2020 tutorial. Link: https://youtu.be/vTRD03_yReI

Keywords: graph neural networks, dynamical models, Gaussian processes, stochastic differential equations, Bayesian methods


Project 27: Deep Quantum Graphical Models

Supervision: Vikas Garg (Aalto University) and Juho Kannala (Aalto University)

Probabilistic graphical models (PGMs) such as Markov Random Fields and Bayesian networks allow us to encode the conditional dependencies between random variables succinctly. However, these models are designed to run on the classical computer. This project is aimed at designing quantum counterparts for PGMs, and implementing them via deep architectures such as graph neural networks [1] that can be executed on a quantum computer. Another goal of the project is to investigate whether it might be possible to simulate these algorithms on a classical computer under some restrictions [2]. We are looking for doctoral students to work at the intersection of quantum machine learning, graphical models, and deep learning. The ideal candidate will have a strong mathematical inclination and training, and excellent programming skills.

References:

Keywords: graphical models, quantum computing, deep learning


Project 28: Atmospheric AI

Supervision: Kai Puolamäki (University of Helsinki); potential co-supervisors Hanna Vehkamäki (University of Helsinki), Leena Järvi (University of Helsinki), Tuomo Nieminen (University of Helsinki)

Artificial intelligence (AI) and machine learning (ML) are making their inroads to atmospheric and earth sciences. There are lots of opportunities to do research in physical sciences more efficiently and to obtain novel results of high impact—both in atmospheric and computer sciences—by developing and applying novel AI methods to solve scientific problems. In this project, we plan to build probabilistic models of measured and simulated natural world phenomena, trained by using simulator outputs or real-world observations, which allow us for example replace computationally expensive simulator runs with faster ML computations, to fill in missing data from observations, and to better understand complex systems and processes and underlying causal relations. Our objective is to also model the interactive data analysis and model building process of the substance area experts (here atmospheric scientists), which allows us to address problems such as how to design the exploratory data analysis workflows and systems and how to best incorporate the knowledge and insights of the experts into the model building process. We are looking for an atmospheric scientist with interest in AI, or a computer scientist who wants to develop AI methodology and work with physics-related applications. We can adjust the work plan and the supervision arrangement depending on the qualifications and interests of the hired person.

Keywords: Atmospheric and earth sciences; exploratory data analysis; automatic experimental design; interactive user modelling; causal inference


Project 29: Neuroadaptive interfaces for generative brain-computer design

Supervision: Tuukka Ruotsalo (University of Helsinki) and Jaakko Lehtinen (Aalto University).

The project aims to establish the scientific foundations for generative brain-computer design. The project combines 1) human preference and critique estimation directly from the human brain manifested as implicit natural human reactions evoked in response to generative model 2) use critique as input for generative models producing that information to a) derive objectives for learning generative models, b) to adjust generative model outputs, and c) develop self-supervised brain-computer interfaces for 3) interactive design. The work is based on our fundamental research on novel neuroadaptive brain-computer interfaces [1,2,3,4] and is conducted in close collaboration with neuroscientists.

The PhD student will develop new types of deep learning methods that can jointly decode human brain responses and associate the decoding process to adapt and learn generative latent models that produce stimuli information evoking the brain responses. and evaluate and test the models in human-computer and brain-computer interaction settings. The student should have experience in deep learning and the associated software development with TensorFlow or similar frameworks. A basic knowledge, or interest to learn, cognitive neuroscience, especially EEG and fNIRS brain imaging methods, are an advantage.

  • [1] Kangassalo Lauri; Spapé, Michiel; Ruotsalo, Tuukka. Neuroadaptive modelling for generating images matching perceptual categories. Scientific reports (Nature), 2020, 10.1: 1-10.
  • [2] de la Torre-Ortiz, Carlos, et al. Brain Relevance Feedback for Interactive Image Generation. In: Proceedings of the 33rd Annual ACM Symposium on User Interface Software and Technology. 2020. p. 1060-1070.
  • [3] Kangassalo, L., Spapé, M., Ravaja, N., & Ruotsalo, T. information gain modulates brain activity evoked by reading. Scientific reports (Nature), 2020, 10.1, 1-10.
  • [4] Davis III, Keith M., et al. Brainsourcing: Crowdsourcing Recognition Tasks via Collaborative Brain-Computer Interfacing. In: Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems. 2020. p. 1-14.

More information about related research is available at: https://www.cs.helsinki.fi/group/intercom/. The project also involves cooperation with University of Copenhagen.

Keywords: brain-computer interfaces, brain feedback, generative adversarial neural networks, deep autoencoders, design


Project 30: Interactive reward elicitation

Supervision: Ville Kyrki (Aalto University), Simo Särkkä (Aalto University), Joni Pajarinen (Aalto University), Antti Oulasvirta (Aalto University), Samuel Kaski (Aalto University)

A major problem in RL is to determine suitable reward functions. Reward design often requires a significant amount of trial-and-error even for experts with experience. Our core idea is to include a model of the user in the reward elicitation process such that the process can take into account the user’s limitations in addition to maximizing information gain, thus moving beyond the noisy oracle paradigm. Moreover, we plan to integrate behavioural (explicit feedback from user) as well as implicit physiological measurements.

Keywords: Reinforcement learning, reward design, reward elicitation, interactive AI, inverse RL


Project 31: Graph based world models for sample efficient and human friendly reinforcement learning

Supervision: Joni Pajarinen (Aalto University), Alexander Ilin (Aalto University), Juho Kannala (Aalto University)

Reinforcement learning has shown promise in computer game play and robotics but learning long-term behavior directly from visual input has been limited to simple tasks and actual task specification has been hard for regular users. Meanwhile deep learning has been able to infer semantic information about objects and their dependencies from visual input in the form of object graphs. To make long-term planning more efficient we will use graphs as dynamic states in reinforcement learning. We will learn models of how the state of the system, that is, the graph changes on agent actions. Using the learned world model together with a reinforcement learning policy representation such as a graph neural network allows the system to generalize over different world scenes. Moreover, the approach allows for measuring closeness of the current scene to the desired one in the form of graph similarity measures. Thus, a non-expert user can specify the task objective in terms of visual scenes, for example, by uploading a set of photos of the desired end state or of states that the algorithm should avoid which are then converted to graphs. We provide good opportunities for applying the methods in mobile robotic manipulation and autonomous driving.

Keywords: Reinforcement learning, model learning, planning, computer vision, decision-making, human feedback, robotics


Project 32: Planning to learn world-models to plan and learn

Supervision: Joni Pajarinen (Aalto University), Alexander Ilin (Aalto University), Simo Särkkä (Aalto University), Ville Kyrki (Aalto University)

Planning to reach long-term goals has allowed for super-human performance in tasks with accurate models [1] while learning has allowed for solving tasks with complex inputs and outputs [2]. However, tasks with complex inputs and outputs that require long term exploration and planning are out of reach for current methods. We will go further and integrate learning and planning. We will use meta-learning [3] to learn a world model and a policy that mixes planning and reinforcement learning such that planning is used for problem parts that require principled exploration and learning is used for parts that require pattern recognition. We will develop methods that integrate learned dynamics models and planning with both model-free and model-based reinforcement learning. We expect the developed methods to enable efficient long-term decision making in high-dimensional continuous and discrete control tasks. We provide good opportunities for applying the methods in mobile robotic manipulation and autonomous driving.

  • [1] Silver, D., Schrittwieser, J., Simonyan, K., Antonoglou, I., Huang, A., Guez, A., Hubert, T., Baker, L., Lai, M., Bolton, A. and Chen, Y., 2017. Mastering the game of Go without human knowledge. Nature, 550(7676), pp.354-359.
  • [2] Hessel, M., Modayil, J., van Hasselt, H., Schaul, T., Ostrovski, G., Dabney, W., Horgan, D., Piot, B., Azar, M.G. and Silver, D., 2018. Rainbow: Combining Improvements in Deep Reinforcement Learning. In AAAI.
  • [3] Finn, C., Abbeel, P. and Levine, S., 2017. Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks. In ICML.

Keywords: Reinforcement learning, machine learning, learning dynamics, control, planning, decision-making


Project 33: Transferable hierarchical reinforcement learning

Supervision: Joni Pajarinen (Aalto University), Alexander Ilin (Aalto University)

We tackle the general problem of learning to make high value decisions in dynamic systems at different time-scales. Major challenges of current approaches include high sample complexity, problem specificity of learned solutions, and opaquity of models. We will develop efficient hierarchical decision making models that allow 1) learning parts of the model separately for efficiency, 2) transferring the model to novel situations including transfer from simulation to the real world, and 3) demonstrating model parts and their interplay to users to bridge the human-computer gap. Contrary to prior work that does not learn models or learns only low level models we learn models on all levels allowing for quick transfer to new tasks and interpretability for users. We are looking for a researcher interested in developing the new decision making models and methods, with options on applying the techniques to autonomous driving and robotics.

Keywords: Reinforcement learning, learning dynamics, control, machine learning, planning, decision making under uncertainty


Project 34: Context based curriculum learning for safe exploration

Supervision: Joni Pajarinen (Aalto University), Ville Kyrki (Aalto University)

Contrary to state-of-the-art curriculum learning that incrementally adapts the underlying task to make learning more efficient [1], we will use curriculum learning to also improve safety in exploration on real systems. We will use context based curriculum learning where the agent controls the context. By controlling the context the agent can explore different behaviors without risking the system or users. In human-robot interaction, the context can be defined in terms of the user, for example, user pose or pose of an object the user is holding. To change the context, the robot asks the human to execute the desired pose or hand out a specific object. With this approach we expect to be able to learn behavior safely on challenging physical systems. Technically the approach will require probabilistic modeling of a safety index (for example, safety margin distance between user and robot) that is learned interactively with a user that controls the context. Then exploration in reinforcement learning will need to integrate the safety index such that a minimum required safety is guaranteed, which can be achieved by controlling the learning curriculum.

[1] Klink, P., D’Eramo, C., Peters, J. and Pajarinen, J., 2020. Self-Paced Deep Reinforcement Learning. Advances in Neural Information Processing Systems (NeurIPS)

Keywords: Reinforcement learning, curriculum learning, safety, planning, decision making under uncertainty


Project 35: Data augmentation, noise and active learning — A Bayesian brain approach

Supervision: Aapo Hyvärinen (University of Helsinki) and Luigi Acerbi (University of Helsinki)

Data augmentation is crucial for modern deep learning methods. However, it is usually done in a very heuristic and manually designed way. Recently, the possibility of learning to do such data augmentation from data, especially in a probabilistic framework, has received increasing attention. Interestingly, something akin to data augmentation occurs naturally in biological brains, which rely on noisy sensory inputs and noisy neural circuits. Biological systems learn to push such noise in directions that leave the underlying inference invariant, effectively learning the local equivariance structure of the inputs. Another important strategy adopted by biological systems is active learning. Here, our goal is to develop probabilistic models which enable learning of data augmentation, in particular by drawing inspiration from Bayesian approaches to brain function, and exploit such probabilistic representations to perform active learning. This project aims at making existing machine learning algorithms more efficient while at the same time elucidating deep connections between data augmentation, noise and Bayesian active learning in neural networks, both artificial and biological. The candidate should have an MSc degree containing a lot of mathematics.

Keywords: Probabilistic modelling, Bayesian brain, Data augmentation, Active learning


Project 36: Bayesian deep active learning for amortized inference of simulator models

Supervision: Luigi Acerbi (University of Helsinki); Jukka Corander (University of Helsinki), Samuel Kaski (Aalto University)

Recent approaches to inference in simulator models exploit the power of flexible deep neural density estimators to iteratively learn a direct mapping from summary statistics of the data to the posterior distribution, skipping the intermediate steps of approximate inference. In some limited cases, the trained networks can be immediately used on new data, achieving the holy grail of amortized Bayesian inference — inference at virtually no cost at runtime. However, training these networks requires a very large number of samples from the model, and the mapping to the posterior has no notion of uncertainty, meaning that the network could fail silently in unseen regions of parameter space. This project is concerned with applying Bayesian principles of uncertainty estimation and active learning to develop a new generation of algorithms for sample-efficient training of robust, safe emulator networks for simulator-based inference. The ideal candidate has prior experience with deep learning and Bayesian methods.

Keywords: Simulator-based inference; Bayesian deep learning; active learning; neural density estimators; amortized inference


Project 37: Fast active-sampling approximate Bayesian inference for everyone

Supervision: Luigi Acerbi (University of Helsinki); Aki Vehtari (Aalto University), Samuel Kaski (Aalto University), Arto Klami (University of Helsinki)

In recent years, a new approach to approximate Bayesian inference has emerged, on the side of the traditional workhorses (MCMC and variational inference). Active-sampling Bayesian inference aims to build posterior distributions (and approximations of the model evidence) in a sample-efficient way, by constructing a statistical surrogate of the posterior (or likelihood), such as via a Gaussian process, and then actively evaluating the log-likelihood or log-joint distribution where needed to efficiently update the surrogate model [1-3]. This approach is similar to Bayesian optimization, but the goal differs in that the goal is to learn the posterior distribution (and/or the marginal likelihood). Crucially, thanks to recent advances in Bayesian nonparametrics, this approach is not limited anymore to “expensive” models, but it could become part of the standard Bayesian workflow for many models, affording calculation of cheap, uncertainty-aware posterior approximations with only a small number of evaluations. This project is concerned with pushing the state-of-the-art of active-sampling Bayesian inference algorithms, in terms of both theory and implementation, to obtain a new instrument for approximate inference which would be widely accessible, fast and failsafe. The ideal candidate has prior experience with Gaussian processes and active learning (e.g., Bayesian optimization), both in theory and with modern software implementations (e.g., GPyTorch).

  • [1] Acerbi L (2018). Variational Bayesian Monte Carlo, NeurIPS.
  • [2] Järvenpää M., Gutmann MU, Vehtari A., and Marttinen P (2020). Parallel Gaussian process surrogate Bayesian inference with noisy likelihood evaluations. Bayesian Analysis.
  • [3] Acerbi L (2020). Variational Bayesian Monte Carlo with Noisy Likelihoods, NeurIPS.

Keywords: Bayesian inference; active learning; Gaussian processes; Bayesian optimization


Project 38: Prior constraints in probabilistic programming

Supervision: Arto Klami (University of Helsinki); Aki Vehtari (Aalto University)

Using prior domain knowledge on model parameters and transformations is at the core of Bayesian modeling, helping to build more interpretable models that can be estimated from less data. This project develops richer ways of encoding prior knowledge, focusing on incorporating (soft and hard) constraints into probabilistic programs. The current tools support simple constraints (non-negativity of a parameters, or linearity of a function) but often the prior knowledge is in form of more complex constraints (e.g. monotonicity or near-linearity of a function, or permutation invariance) that remain challenging. Building on existing theoretical foundations for specific cases, you will work on developing both the theory and practical inference algorithms for handling such constraints.

The position is ideal for candidates with strong background in Bayesian modeling or machine learning. Your main task is to develop the required computational methods and ideally proceed to implement them into existing probabilistic programming tools in collaboration with others. We already have several concrete applications with such prior knowledge (e.g. physical knowledge in material design and cognitive theories in decision-making), and you will work in collaboration with other FCAI projects to apply the methods in selected interesting use-cases.

Keywords: Probabilistic programming, Bayesian modeling, Prior knowledge


Project 39: Visualization in modeling workflow

Supervision: Aki Vehtari (Aalto University); Antti Oulasvirta (Aalto University); Arto Klami (University of Helsinki)

We develop AI techniques needed for systems which can help their users make better decisions and design better solutions across a range of tasks from personalized medicine to materials design. A core insight in developing such AIs is that they need to have world models for understanding the world and interacting with it, and user models for understanding the user and interacting with them. Many parts of the probabilistic modeling workflow benefit from visualization. This project develops tools for AI-assisted visualizations using AI which has a theory of mind of the user. The work will be built on existing theory in cognitive sciences and human-computer interaction. The goal is to generate visually appealing, task-specific, and informative visualizations with controllable complexity depending on the amount of information that is available, required, and sensible given the expertise of the user.

Keywords: Interactive probabilistic modeling, modeling workflow, visualization, uncertainty quantification, decision-making


Project 40: Computer assisted Bayesian workflow

Supervision: Aki Vehtari (Aalto University); Arto Klami (University of Helsinki); Antti Oulasvirta (Aalto University)

Probabilistic programming languages make it easier to specify and fit Bayesian models, but this still leaves us with many options regarding constructing, evaluating, and using these models, along with many remaining challenges in computation. Using Bayesian inference to solve real-world problems requires not only statistical skills, subject matter knowledge, and programming, but also awareness of the decisions made in the process of data analysis. All of these aspects can be understood as part of a tangled workflow of applied Bayesian statistics. Beyond inference, the workflow also includes iterative model building, model checking, validation and troubleshooting of computational problems, model understanding, and model comparison. The goal is to develop for different parts of the workflow self-diagnosing tools that can be used as part of interactive computer assisted workflow for probabilistic model building and Bayesian data analysis.

[1] Gelman, Vehtari, Simpson et al (2020). Bayesian workflow. https://arxiv.org/abs/2011.01808

Keywords: probabilistic modeling workflow, Bayesian workflow, diagnostics