Department of Computer Science

Summer jobs for 2021

PML Research group / photo Matti Ahlgren

The Department of Computer Science is now 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 some other university to work with us during the summer 2021. 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 and doctoral students & postdocs at the department.

 

Ready to apply?

See the complete list of the available topics here below, choose the topic(s) that interest you the most and list them on the application form.

Please submit your application through our recruitment system by 17 January 2021 you will find the link to the application here: http://www.aalto.fi/en/about/careers/jobs/view/3110/

To apply, please share the following application materials with us:

  1. Motivation letter and CV (in one single pdf-file)
  2. Up-to-date transcript of records (unofficial is ok)

 

On the application form:

  1. List the number of the topics that you find most interesting in the application section in order of preference in "Number(s) of topic(s) you are interested in".
  2. Click the "Preview" button to see your application and then click "Submit"

 

Are you an international student or coming from abroad?

Please check the Aalto Science Institute AScI internship programme for international summer employees.
AScI arranges activities for international summer employees who have applied through their call and helps in finding the apartment in Espoo.

More information:

If you have questions regarding the applying please contact HR Secretary Sanni Kirmanen [email protected]

About the department

World-class research and education in modern computer science to foster future science, engineering and society. The work combines fundamental research with innovative applications.

Read more
Computer Science building / Tietotekniikan talo / photo Aalto University, Matti Ahlgren

1. Statistical or psychological theories for user security

Professor/Academic contact person for further information on topic: Prof. Janne Lindqvist

Contact info: [email protected]

Aalto Human-Computer Interaction and Security Engineering Lab http://lindqvistlab.org at the Aalto Department of Computer Science is recruiting a summer intern. We will be recruiting a student to contribute to our research on the application of statistical theory or psychological theory to the study of user security. For example, we will study the cognitive mechanisms responsible for password usability and the psychological principles driving security decisions. Our project will engage the summer intern in every step of the research process including research design, statistical analysis and provide the intern with a unique exposure to the application of psychology and statistics in human-computer interaction and usable security. Projects may also include Bayesian data analysis. Accepted student also has the opportunity to participate to activities of Helsinki-Aalto Institute for Cybersecurity (HAIC).

____

 

2. Statistical or psychological theories for combatting computer fatigue and multitasking

Professor/Academic contact person for further information on topic: Prof. Janne Lindqvist

Contact info: [email protected]   

Aalto Human-Computer Interaction and Security Engineering Lab http://lindqvistlab.org at the Aalto Department of Computer Science is recruiting a summer intern. We will be recruiting a student to contribute to our research on the application of statistical theory or psychological theory to the study of combatting computer fatigue and issues with multitasking. For example, we will study the cognitive mechanisms responsible for computer fatigue and psychological principles related to issues with multitasking. Our project will engage the summer intern in every step of the research process including research design, statistical analysis and provide the intern with a unique exposure to the application of psychology and statistics in human-computer interaction. Projects may also include Bayesian data analysis.

____

 

3. Online learning material development

Professor/Academic contact person for further information on topic: Dr. Ari Korhonen
Contact info: [email protected]

We are hiring summer interns to develop online course materials for CS-A1141/CS-A1143 Data Structures and Algorithms Y (Tietorakenteet ja algoritmit Y) and CS-A1130 Applications of Computing (Tietotekniikka sovelluksissa). We expect the candidates to have passed at least one of the courses above and have good Python programming skills. Teaching Assistant experience is considered to be an advantage.

____

 

4. Investigating adversarial attacks

Professor/Academic contact person for further information on topic: Prof. Antti Ylä-Jääski
Contact info: [email protected]

Adversarial attacks are attacks that cause misclassification in neural network classifiers, but not in humans for example. Since neural classifiers are increasingly being deployed in safety-critical systems, it is important to understand these attacks better in order to defend against them. We aim to investigate adversarial attacks.

Previous experience with python, deep learning and Keras required. Previous experience with adversarial machine learning preferred. Curiosity encouraged.

____

 

5. Deep Learning methods for extreme-scale classification

Professor/Academic contact person for further information on topic: Prof. Rohit Babbar
Contact info: [email protected]

Automatic classification of data to hundreds of thousand labels are common in Machine learning problems such as ranking, recommendation systems and next word prediction. Apart from the computational problem of scalability, data scarcity for individual labels poses a statistical challenge and especially so for data hungry deep methods. The goal of the project is to investigate deep learning based architectures and adapting the well known techniques such as Attention mechanism for simultaneously addressing the computational and statistical challenge in learning with large output spaces. As the target domain is textual data, the project also involves exploring recent advances in NLP, such as Bert and TransformerXL, towards exploring the common grounds for further research in this area.

____

 

6. Harjoitustehtävien suunnittelu ja toteuttaminen kurssille CS-A1111 Ohjelmoinnin peruskurssi Y1

Professor/Academic contact person for further information on topic: Dr. Kerttu Pollari-Malmi
Contact info: [email protected]

Kesätyöntekijän tehtävänä on suunnitella uusia harjoitustehtäviä kurssille CS-A1111 Ohjelmoinnin peruskurssi Y1 sekä toteuttaa laadituille tehtäville automaattinen arvostelu A+-järjestelmässä. Haen tehtävään teekkaria, joka keksii helposti uusia tehtäväideoita. Hyvä ohjelmointitaito on myös tarpeen. Koska kurssin kieli on suomi, erinomainen suomen kielen taito on välttämätön.

Harjoitustehtävien ja niiden tarkistimien laatimisen lisäksi kesätyöntekijä toimii lisäassistenttina kesäkurssin harjoitustehtäväneuvonnassa muutaman tunnin viikossa.

___

 

 

7. Simulation and analysis of dynamical systems

Professor/Academic contact person for further information on topic: Dr. Riku Linna

Contact info: [email protected]

The summer intern will participate in implementing algorithms for studying nonlinear dynamics and chaotic systems. The work will comprise implementing a simulated dynamical system for  generation of time series data and implementation and usage of a machine learning method, namely reservoir computing, to analyse the system based on this generated data. The summer intern will implement a well defined new part that will be added to the algorithms that are already in use within the group. The motivation is to develop methods for understanding nonlinear dynamical systems based on a presumed system model and measured data. Such methods are widely applicable in various fields, such as finance, involving prediction based on time-series data. A prominent potential application area is the data-driven analysis of connectivity underlying nonlinear dynamics of (stimulated) activation in the brain.

The student should have a fair command of programming in python and an interest and general understanding of at least one of the following fields: machine learning methods, dynamical systems, time-series forecasting, and stochastic processes. The purpose of summer internship is also to learn and get acquainted with the research field, so expertise in the named fields is not necessary.

____

 

8. Pääassistentti kesäkurssille CS-A1111 Ohjelmoinnin peruskurssi Y1

Professor/Academic contact person for further information on topic: Dr. Kerttu Pollari-Malmi
Contact info: [email protected]

Pääassistentin tehtäviin kuuluu laatia harjoitustehtävät kesäkurssille (osin vanhojen tehtävien pohjalta, osin uusia), toteuttaa tehtävien automaattinen tarkistus A+-järjestelmään, neuvoa kesäkurssin opiskelijoita joko paikan päällä järjestettävissä harjoitusryhmissä tai etäyhteyden avulla (sen mukaan, voidaanko harjoitusryhmiä kesällä järjestää kampuksella) sekä vastata opiskelijoiden kysymyksiin kurssin www-neuvontakanavassa (Piazza tai vastaava).

Työt pitää aloittaa osa-aikaisena joko huhti- tai toukokuun alussa (työntekijän valinnan mukaan), jotta harjoitustehtäviä olisi tarpeeksi valmiina ennen kuin kurssi alkaa kesäkuun alussa. Työ päättyy 20.8. Tehtävästä maksetaan yhteensä kolmen kuukauden täysiaikaista työtä vastaava palkka.

Työ vaatii hyvää ohjelmointitaitoa sekä Python-ohjelmointikielen osaamista. Aiempi kokemus kurssiapulaisen tehtävistä tai muu opetuskokemus on etu. Koska kurssin kielenä on suomi, työ vaatii erinomaista suomen kielen taitoa.

____

 

9. Ohjelmoinnin peruskurssi Y2:n kehittäminen

Professor/Academic contact person for further information on topic: Sanna Suoranta
Contact info: [email protected]

Etsin kahta kesäteekkaria kehittämään tietotekniikan sivuaineen kurssia CS-A1121 Ohjelmoinnin peruskurssi Y2 ja sen englanninkielistä sisarkurssia. Kesäteekkarien tehtävä on päivittää kurssin harjoitustehtäviä, joiden aiheena ovat olio-ohjelmointi Python-kielellä, yksikkötestit, UML-suunnittelu, GIT sekä käyttöliittymäohjelmointi PyQt-kirjaston avulla. Lisäksi kesäteekkarit päivittävät vanhoja ja suunnittelevat uusia projektiaiheita kurssille. Tämä kesäteekkarin työ edellyttää suomen, englannin ja python-ohjelmointikielen hyvää hallitsemista sekä hyvin suoritettua Ohjelmoinnin peruskurssi Y2:sta.

____

 

10. Training probabilistic circuits on large scale datasets

Professor/Academic contact person for further information on topic: Dr. Martin Trapp / Prof. Arno Solin
Contact info: [email protected]

Probabilistic circuits, such as sum-product networks and probabilistic sentential decision diagrams, are a promising class of deep probabilistic models that unify ideas from classical AI and modern deep learning. In contrast to many other probabilistic modelling families, probabilistic circuits guarantee exact and efficient computation of many probabilistic inference tasks. However, training a circuit on massive datasets is a challenging task due to their sparse architecture. The aim of this project is to adopt work on coresets to enable efficient parameter learning in large scale settings. The prospective candidates should be familiar with topics related to probabilistic machine learning and should ideally have experience with python deep learning frameworks.

References:

  1. Peharz et al. (2020). Einsum Networks: Fast and Scalable Learning of Tractable Probabilistic Circuits, in ICML. http://proceedings.mlr.press/v119/peharz20a/peharz20a.pdf
  2. Lucic et al. (2018). Training Gaussian Mixture Models at Scale via Coresets, in JMLR. https://www.jmlr.org/papers/volume18/15-506/15-506.pdf

____

 

 

11. Towards Integrable Continuous Normalizing Flows

Professor/Academic contact person for further information on topic: Dr. Martin Trapp / Prof. Arno Solin
Contact info: [email protected]

Normalizing flows use bijective, differentiable transformations to represent complex probability distributions. Recent work has focused on so-called continuous normalizing flows, which utilize a formulation in form of an ordinary differential equation (ODE) to represent more flexible flow architectures. However, even though flow type models allow tractable computation of densities by construction they lack the ability to compute more complex probabilistic inference tasks, e.g. marginals. In parallel, there has been an increasing interest in probabilistic circuits, which use a much more restricted family of neural network architectures to guarantee exact computation of arbitrary marginals, conditions and moments. The aim of this project is to explore ways to utilize ideas from probabilistic circuits to enable exact computation of marginals in continuous normalising flow models. The prospective candidates should be familiar with topics related to probabilistic machine learning and should ideally have experience with the Julia programming language or alternatively experience with Python frameworks.

References:

  1. Pevny et al. (2020). Sum-Product-Transform Networks: Exploiting Symmetries using Invertible Transformations, in PGM. https://pgm2020.cs.aau.dk/wp-content/uploads/2020/09/pevny20.pdf
  2. Z. Kolter, D. Duvenaud, and M. Johnson (2020). Deep Implicit Layers Tutorial at NeurIPS. http://implicit-layers-tutorial.org

____

 

 

12. Efficient Model-based Path Integral Stochastic Control

Professor/Academic contact person for further information on topic: Dr. Vincent Adam / Prof. Arno Solin
Contact info: [email protected]

Stochastic optimal control (SOC) is the problem of controlling a known noisy nonlinear dynamical system, modeled by a stochastic differential equation (SDE) [1]. SOC has applications in many domains including robotics and underlies many state-of-the-art reinforcement learning algorithms. When the system is unknown or non-stationary, its statistics and parameters can be learned from experience. Gaussian processes (GPs) can be used to incorporate a priori assumptions about the system (e.g., the drift of the SDE) into the problem, leading to sample-efficient algorithms, i.e. algorithms that can achieve good performance with a small number of interactions with the system. However, GPs suffer from poor computational scaling making such algorithms either prohibitively slow and unfit to real world settings or require approximations that impair their performance [2]. Recent advances in GP approximations now provide computationally efficient ways to learn and sample from GPs with guarantees [3].

The aim of this project would be to combine the sample based Path-integral formulation of SOC and these recent advances in GP inference and sampling to derive efficient model based SOC algorithms. The main aim of the project would be to derive, implement and evaluate the resulting algorithm. If time allows, a more theoretical analysis of the algorithm could be done. The prospective candidates should be familiar with topics related to probabilistic machine learning and/or reinforcement learning and should ideally have experience with Python.

References:

  1. Kappen (2008). Stochastic optimal control theory, ICML tutorial
  2. Pan et al. (2014). Model-based Path Integral Stochastic Control: A Bayesian Nonparametric Approach, https://arxiv.org/abs/1412.3038
  3. Wilson et al. (2020). Pathwise Conditioning of Gaussian Processes,  https://arxiv.org/abs/2011.04026

____

 

 

13. Incorporating Physics into Deep Learning with Uncertainty

Professor/Academic contact person for further information on topic: Dr. Vincent Adam / Prof. Arno Solin
Contact info: [email protected]

Dynamical systems are governed by the laws of physics. Learning a dynamical system supporting accurate long-term predictions from observations of trajectories can be greatly enhanced by incorporating the right inductive biases from physics (e.g., in the form of constraints and conservation laws). Recent work has successfully incorporated laws of physics into deep learning [1, 2]. For example, neural networks were used to parameterize the parameters of the Hamiltonian of a system and learning was achieved through back propagating through the solution of the equations of motion derived from this Hamiltonian [3]. The purpose of this project would be to investigate how to best incorporate uncertainty about parameters in such models. Parameter uncertainty leads to uncertain future predictions which are crucial in control scenarios where driving the system in some particular configurations is very costly. The prospective candidates should be familiar with topics related to classical physics, probabilistic machine learning or deep learning and should ideally have experience with deep learning frameworks such as TensorFlow.

References:

  1. Lutter et al. (2020). Deep Lagrangian Networks: Using Physics as Model Prior for Deep Learning, https://arxiv.org/abs/1907.04490
  2. Steindor Saemundsson et al. (2020). Variational Integrator Networks for Physically Structured Embeddings, ICML.
  3. Chen et al. (2018). Neural Ordinary Differential Equations, NeurIPS.

____

 

 

14. Non-stationary temporal priors for large-scale modelling

Professor/Academic contact person for further information on topic: Dr. William Wilkinson / Prof. Arno Solin
Contact info: [email protected]

We are looking for motivated, skilled, and open-minded summer students with an interest in real-time inference and application of probabilistic machine learning methods to practical applications. This project is concerned with non-stationary Gaussian process models and stochastic differential equations for temporal and spatio-temporal data, with possible applications in air pollution modelling, sensor fusion, or computer vision, where the problems are inherently noisy and uncertainty quantification plays an important role. An important part of the project is implementing some of these models in JAX. Successful applicants are expected to have an outstanding record in computer science, mathematics, statistics, or a related field, and familiarity with some of the topics mentioned above.

References:

  1. Arno Solin (2020). Machine Learning with Signal Processing. ICML tutorial. Link: https://youtu.be/vTRD03_yReI
  2. William J. Wilkinson, et al. (2020). State Space Expectation Propagation: Efficient Inference Schemes for Temporal Gaussian Processes. In ICML. Pre-print: https://arxiv.org/abs/2007.05994

____

 

 

15. WebXR based AR indoor navigation client

Professor/Academic contact person for further information on topic: Prof Antti Ylä-Jääski    
Contact info: [email protected]

The WebXR Device API provides access to input (pose information from headsets, eyeglasses, or mobile phones) and output (hardware display) capabilities commonly associated with Virtual Reality (VR) and Augmented Reality (AR) devices. It allows you to develop and host VR and AR experiences on the web.

We have earlier developed a native AR indoor navigation application on Android. It utilizes ARCore and our native C++ libraries to calculate the user’s position and facing direction in indoor environments. In this project, you will make use of the WebXR device API to port this AR navigation application to the WEB browser. You will use JavaScript for the WEB browser development and WebAssembly to link our native C++-libraries into the system. You will also get experience in ARCore since our navigation application makes use of it.

____

 

 

16. Sample-efficient deep (reinforcement) learning  

Professor/Academic contact person for further information on topic: Prof. Alexander Ilin
Contact info: [email protected]
     
Our research group works on developing learning systems capable of generalizing to new concepts and tasks from a small number of demonstrations during training. This kind of a sample-efficient generalization is a must for digital assistants that are truly cooperative and responsive, operating in heterogenous real world environments. Deep learning is a powerful learning tool which currently requires a large amount of training data and shows limited ability to generalize to out-of-distribution data. We aim at improving sample efficiency and generalization abilities of deep learning systems. To improve sample-efficiency, we enhance the learning process by introducing relevant auxiliary tasks. In supervised learning, auxiliary tasks come from modeling unlabeled data (for example, via self-supervision or contrastive learning). In reinforcement learning, the auxiliary tasks come from building a predictive model of the environment. To improve the generalization abilities, we develop models with inductive biases that facilitate better transfer of knowledge and faster adaptation to new tasks. We develop world models that consist of a hierarchy of re-usable components (for example, hierarchical reinforcement learning) and build object-based representations of the world. The developed object-centric models should generalize well to different combinations and visual appearances of interacting objects, thus increasing the generalization ability. Object-centric models also improve the interpretability of the machine learning system as humans are used to reason about the world in terms of objects and their interactions. The developed models combine classical deep learning architectures (such as convolutional neural networks) and neural networks with the relational inductive bias (such as graph neural networks and transformers). In this project, you will work on one of the topics related to this theme.

____

 

 

17. Designing for better Developer Experience: Web + mobile research tool design and development

Professor/Academic contact person for further information on topic: Dr. Fabian Fagerholm
Contact info: [email protected]

We are building a web and mobile application for performing and presenting research on developer experience. We seek students with interest in developer experience research, web development, mobile development, user interface design, and/or data visualisation.

Developer experience refers to the cognitive, motivational, and affective experience that software developers have while developing software. We have previously collected and analysed research related to cognition in software development and are now organising the material into web site form. We are also building tools for collecting experience data from software developers.

Your task could focus on 1) web and mobile development, 2) user interface and visual design, or 3) research instrument development, or a combination, depending on your skills and interests. You would work in a small team that plans, designs, and implements web and mobile components. The work can also include research-related tasks.

Required background and skills (labelled according to focus):

  • Full-stack web app development skills; both static and dynamic web technologies; Node.js, Python, or similar. (Development)
  • Mobile app development skills; Android, iOS, Flutter considered a plus. (Development)
  • Good software development practices; version control (git), automated testing, ability to write maintainable code. (Development)
  • Knowledge and skill in visual and user interface design. (User interface)
  • Ability to turn user interface sketches into compliant web pages; HTML5, CSS. (User interface)
  • Knowledge of user research methods; quantitative (statistical, machine learning) and qualitative data analysis. (Instrument development).
  • Knowledge of user data collection; surveys, observation, log data analysis, non-intrusive behaviour measurement. (Instrument development)
  • Ability to analyse, organise and visualise data; automated analysis/visualisation for quantitative data. (All)
  • Understanding of accessibility, security, and data protection issues. (All)

____

 

 

18. Modern ubiquitous applications: from devices to the cloud

Professor/Academic contact person for further information on topic: Prof. Mario Di Francesco
Contact info: [email protected]

Modern applications that are ubiquitous – namely, everywhere – rely on two key components. First, on embedded devices such as mobile phones, wearables and smart objects in the Internet of Things to interact with users and collect information from the physical environment. Second, on a cloud or edge computing infrastructure to support different types of applications requiring a substantial amount of processing and storage, such as those involving machine learning. The major challenges in realizing such applications include efficient resource utilization at both devices and the supporting infrastructure, reliability, and user friendliness. The goal of this project is to investigate some of these aspects in the context of the research carried out in our research group. See also https://users.aalto.fi/~difram1/ for additional details.

Required skills: experience with Android application development or embedded systems programming, solid understanding of data analysis and (or) machine learning techniques.

Desired skills: some knowledge on human computer interaction and (or) computer vision, familiarity with cloud and web technologies.

____

 

 

19. Machine Learning for Health (ML4H)

Professor/Academic contact person for further information on topic: Prof. Pekka Marttinen
Contact info: [email protected]

Recent years have witnessed an accumulation of massive amounts of health related data, enabling researchers to address diverse problems such as: how to allocate healthcare resources fairly and efficiently, how to provide personalized guidance and treatment to users based on real-time data from wearable self-monitoring devices, or how to use genomic data to understand disease or antibiotic resistance. Central challenges in ML4H include the massive amount of diverse data from multiple data sources, going beyond correlation to learn about causal relations between relevant variables, interpreting the models, and assessing the uncertainty of predictions, to name a few. We tackle these by developing new models and algorithms which leverage modern principles of machine learning, combining techniques such as deep neural networks, probabilistic methods, interactive machine learning, attention, and generative models. Examples of our ongoing interdisciplinary projects include: analysis of nationwide healthcare register data, mobile health, genomics, antibiotic resistance, and epidemiology. We are looking for summer interns with an outstanding study record in computer science, statistics, applied mathematics, or a related field, and a passion to put these skills to use in an interdisciplinary research project to address some of the most burning challenges in today’s society.

____

 

 

20. Foveated Rendering Techniques - A comparison

Professor/Academic contact person for further information on topic: Prof. Antti Ylä-Jääski
Contact info: [email protected]

Foveated Rendering techniques use more rendering resources where the user is looking in a frame, reducing rendering resources elsewhere in the frame to improve graphics performance. The frame rendered using foveated rendering roughly corresponds to the human visual acuity, particularly accounting for the phenomenon of Foveation.

Various techniques have been developed to implement foveated rendering, the differences ranging from an overhaul of rendering pipeline to modifications in the classic rendering pipeline. An example of foveated rendering implementation is Nvidia's Variable Shading Rate API which supports varying the number of shading passes per x pixels at run time, allowing implementation of foveated rendering. The student's task would be to identify and implement other foveated rendering techniques, preferably as unity plugins for a comparative study in terms of performance and image quality.

____

 

 

21. QoE of Cloud Rendered VR/XR

Professor/Academic contact person for further information on topic: Prof. Antti Ylä-Jääski

Contact info: [email protected]

Rendering Virtual/Cross Reality content is quite resource intensive. For this reason, typical VR head mounted devices (HMD) need (and are tethered to)  a powerful computer with sufficient graphics computing resources to provide a high quality of experience (QoE).

Cloud rendered VR/XR seeks to move the resource intensive rendering operations to a graphics cloud, using video streaming to display the content rendered in the cloud on a thin client VR HMD. However, strict constraints in terms of video quality and latency have to be fulfilled to deliver a sufficiently high QoE of experience. There are various approaches to fulfilling these constraints, some of them, by modifying and optimizing the video streaming pipeline. The intern's task would be to develop a unity application which allows us to measure subjective QoE of VR content. The application would be for example a game, which engages the user in various tasks, each task engaging the user to focus on a particular aspect of the VR experience. This application would then be plugged into our remote rendering set-up. The candidate must have some experience in Unity development.

____

 

 

22. Egocentric action anticipation

Professor/Academic contact person for further information on topic: Dr. Jorma Laaksonen / Selen Pehlivan Tort
Contact info: [email protected]

Egocentric action recognition has been a promising topic in Computer Vision in recent years. Among those efforts on egocentric action studies, some have been devoted to action anticipation in first-person vision. Human action anticipation aims to predict future unseen actions from the observation of preceding video segments. It requires understanding past and guessing user intentions to predict the next action in advance. This would be very helpful in real world applications including surveillance, self-driving, human-robot interaction systems. The study includes the review of the most recent deep learning systems on egocentric (first-person view) action anticipation and proposing a model. We will conduct research on large scale egocentric video datasets.

Prerequisite:

  1. basic knowledge of Computer Vision (image/video processing) and Machine Learning,
  2. familiarity with large scale datasets and
  3. programming skills with deep learning frameworks such as PyTorch. 

____

 

 

23. Massively Parallel Algorithms for Graph Problems

Professor/Academic contact person for further information on topic: Prof. Jara Uitto
Contact info: [email protected]

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.

___

 

24. Causal inference with simulators

Professor/Academic contact person for further information on topic: Prof. Samuel Kaski, Alex Aushev
Contact info: [email protected], [email protected]

Simulators are often considered as black-box generative models, where the relationship between its parameters is unknown. However, knowing causal relationships between variables of the model can significantly improve the inference quality. This is especially the case when the generative model is an interaction between a user and a simulator (experimental design) or when only individual model components are known but not their interaction (multi-agent system). For this intern position you will join the Aalto Probabilistic Machine Learning group to develop causal simulator-based probabilistic models in a likelihood-free inference setting.

___

 

 

25. Task-oriented Active Learning, with application to healthcare and medicine

Professor/Academic contact person for further information on topic: Prof. Samuel Kaski, Louis Filstroff and Iiris Sundin
Contact info: [email protected], [email protected], [email protected]

Active learning is a special case of supervised learning where the algorithm is able to query for additional training examples, often with the help of a human expert. Active learning basically comes down to selecting the most informative unlabeled example. In the scenario where the model is then used for a decision-making task, we aim at developing new active learning techniques which take into account this task.

We are looking for a summer intern to join the Probabilistic Machine Learning group to help us develop these techniques. A notable application of this work is in healthcare and medicine, where such techniques can be applied to help doctors choose the best treatment for a specific patient (so-called personalized medicine).

Students with good programming skills, and with a strong background in machine learning or statistics are encouraged to apply.

____

 

 

26. Likelihood-free Inference for Cognitive Neuroscience

Professor/Academic contact person for further information on topic: Prof. Samuel Kaski, Dr. Nitin Williams
Contact info: [email protected], [email protected]

Exciting modelling efforts in Cognitive Neuroscience have furnished understanding on biophysical mechanisms underlying human Neuroscience data. However, fitting these models to experimental data is hampered by intractability of the model likelihood functions. Likelihood-free inference (LFI) techniques promise to bridge this gap. In this project, we will use LFI methods from Approximate Bayesian Computation (ABC) to fit complex dynamical models to human Neuroscience data, thus yielding valuable insight into mechanisms producing these data. You will gain knowledge of ABC techniques, experience with the ELFI (Engine for Likelihood-Free Inference) toolbox (elfi.ai), an introduction to models in Cognitive Neuroscience, and useful experience in scientific computation.

____

 

 

27. Benchmarking differential privacy

Professor/Academic contact person for further information on topic: Prof. Samuel Kaski, Joonas Jälkö and Lukas Prediger
Contact info: [email protected], [email protected], [email protected]

In recent years differential privacy has been established as the prevailing privacy notion. In practice, differential privacy is often achieved by adding noise into a statistical query to mask out any individual contribution. The amount of noise added affects the privacy guarantees, and needs to be carefully calibrated to introduce as little bias as possible while maintaining the strongest possible theoretical guarantees. We are looking for a SUMMER INTERN to empirically stress test known DP algorithms to explore the gap between theory and practice. A suitable candidate has a strong background in math, especially in probability, and in programming (Python preferred). Join our quest for developing machine learning with strong privacy guarantees!

___

 

 

28. Privacy-preserving machine learning

Professor/Academic contact person for further information on topic: Prof. Samuel Kaski, Joonas Jälkö
Contact info: [email protected], [email protected]

We develop methods for learning from data given the constraint that privacy of the data needs to be preserved. This problem can be formulated in terms of a concept called differential privacy, and we have for instance introduced ways of sharing strongly anonymized versions of data (https://arxiv.org/abs/1912.04439). A couple of “minor” problems still remain; come to solve them with us! Requirements: strong background in math, especially in probability, decent skills in programming, and/or a steep gradient in the learning curve.

____

 

 

29. Learning-dynamics and AI Safety

Professor/Academic contact person for further information on topic: Prof. Samuel Kaski, Mert Çelikok
Contact info: [email protected], [email protected]

A grand goal of AI research is to develop agents that can learn continually and interact with other humans and agents. Such agents can represent their humans’ interests in settings like negotiation and modelling. Clearly, safety and security of such systems are of paramount importance. In this project, we will use the tools of dynamical systems to analyze how different reinforcement learning algorithms learn things over time and adjust their behaviour in safety-critical environments. We will look for patterns such as cycles and fixed-points to analyze how algorithms learn behaviour in such environments.

Keywords: AI Safety, Learning dynamics, human-AI interaction, reinforcement learning"

____

 

 

30. Probabilistic machine learning for genomics and precision medicine

Professor/Academic contact person for further information on topic: Prof. Samuel Kaski, Tianyu Cui
Contact info: [email protected], [email protected]

We are looking for a summer trainee to join us in developing new methods for genomics. Possible projects include: informative priors for improving prediction, finding genetic interactions underlying disease risk, meta-analysis for GWAS. The work involves probabilistic modelling, deep learning and meta learning, and the methods will also be needed and used in precision medicine later.

____

 

 

31. Probabilistic user models for image design

Professor/Academic contact person for further information on topic: Prof. Samuel Kaski, Petrus Mikkola
Contact info: [email protected], [email protected]

Successful machine learning techniques have been developed for generating novel images (GANs) and learning user preferences on relatively high-dimensional spaces (GP based preference learning). What the techniques do not solve successfully yet, is how to combine these techniques to personalize the image design task for the user. We now have new ideas on how to do that based on advanced interactive user modelling, and we would welcome a student excited on this problem to join us solve the challenge. The work requires a strong background in computer science or statistics, and a willingness to learn quickly.

___

 

 

32. Bayesian preference learning and recommender systems

Professor/Academic contact person for further information on topic: Prof. Samuel Kaski, Petrus Mikkola
Contact info: [email protected], [email protected]

The recent Bayesian machine learning techniques, which are build upon psychometrics and the economic utility theory, enable efficient preference learning and belief elicitation. The fundamental idea is to model the user’s latent valuation function (aka utility function) as a Gaussian process. From the application perspective, the research is closely related to recommender systems. The two main types of recommender systems are based on similarity computations over users (collaborative filtering) or over items (content-based filtering). The latter relates to the same problem that can be dealt with Bayesian preference learning. The goal of the summer project is to synthesize the research on recommender systems and Bayesian preference learning into a coherent piece of text, and to conduct an experimental comparison of reviewed methodologies. The work requires a strong background in computer science or statistics, and a willingness to learn quickly.

___

 

 

33. Probabilistic interactive user models for interactive AI

Professor/Academic contact person for further information on topic: Prof. Samuel Kaski, Pierre-Alexandre Murena
Contact info: [email protected], pier[email protected]

Most machine learning systems operate with us humans, to augment our skills and assist us in our tasks. In environments containing human users, or, more generally, intelligent agents with specific goals and plans, the system can only help them reach those goals if it understands them. Since the goals can be tacit and changing, they need to be inferred from observations and interaction. Join us in developing the probabilistic interactive user models and inference techniques needed to understand other agents and how to assist them more efficiently!

Keywords: active learning, experimental design, knowledge elicitation, multi-agent learning, machine teaching, reinforcement learning

https://research.cs.aalto.fi/pml/

https://aaltopml.github.io/machine-teaching-of-active-sequential-learners/

___

 

34. Bayesian Experimental Design for model-based Reinforcement Learning

Professor/Academic contact person for further information on topic: Prof. Samuel Kaski, Mert Çelikok

Contact info: [email protected], [email protected]

In this summer project the goal is to explore Bayesian Experimental Design approaches for controlling the exploration-exploitation trade-off in model-based Reinforcement Learning. Within Reinforcement learning we aim to choose the optimal actions to achieve a task, and Bayesian Experimental Design gives a framework to make these decisions under the model

uncertainty. You will work as part of the PML research group, which has many complementary research projects in model-based Reinforcement Learning and Experimental Design. Requirements: a strong background in statistics or computer science.

https://research.cs.aalto.fi/pml

____

 

 

35. Assorted topics in Gaussian process modeling and particle-based inference

Professor/Academic contact person for further information on topic: Prof. Samuel Kaski, Zheyang Shen
Contact info: [email protected], [email protected]

This summer project focuses on theoretical topics with respect to the modeling and inference of Gaussian process models, and/or particle-based inference methods. Students are free to choose a combination of the above topics, and conduct theoretical studies as well as applications of the theories.

Gaussian processes (GP) can be seen as infinite-width limits of neural networks, with tractable, albeit often not scalable, uncertainty measures. In terms of GP modeling, students can choose to explore the expressiveness of GP models, specifically about kernel choices connected to deep neural networks. In terms of scalable inference for GPs, students can study the scalable approximations of GP models, specifically sparse methods based on inducing points. In addition, study of other topics tangentially related to GPs are also welcomed, such as deep GPs, and the application of neural tangent kernels.

Another possible topics touches upon the issue of extracting samples from intractable probability distributions, namely particle-based inference derived from Wasserstein gradient flow theories. Various forms of Markov chain Monte Carlo (MCMC) methods corresponds to closed-form, albeit partially intractable Wasserstein gradient flows (WGF), allowing for more sample-efficient alternatives to MCMC. In this project, students can study the connection between MCMC and WGF, and the corresponding WGF-based algorithms.

Students having strong background in mathematics and interest in modeling and inference are especially encouraged to apply. Skills and interest in Python programming are considered a plus.

https://research.cs.aalto.fi/pml

____

 

 

36. Bayesian deep learning

Professor/Academic contact person for further information on topic: Prof. Samuel Kaski, Dr. Markus Heinonen
Contact info: [email protected], [email protected]

We are looking for a 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."

_____

 

 

37. Deep learning with differential equations  

Professor/Academic contact person for further information on topic: Dr. Markus Heinonen
Contact info: [email protected]
 
We are looking for an exceptional and motivated 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. Possible topics include interpretable neural ODEs, to modelling data augmentations as flows, and to probabilistic neural ODEs. The work is supported by multiple active research lines at the department. Background with machine learning, statistics or math learning will be useful.

____

 

 

38. Deep generative modeling for personalised medicine

Professor/Academic contact person for further information on topic: Prof. Harri Lähdesmäki
Contact info: [email protected]

We are looking for several summer interns to work on deep generative and probabilistic modeling for biomedical and health applications. Our on-going work in this topic involves several important biomedical challenges, such as (i) analysis of large-scale heterogeneous health data from Finnish biobanks, (ii) personalized prediction of immunotherapy efficiency for cancer patients using e.g. modern single-cell data, (iii) analysis and design of protein sequences (including Sars-Cov-2), and (iv) longitudinal analysis of multi-omics data from biomedical studies. We are developing novel deep generative modeling methods to analyse aforementioned health data. Deep learning methods include variational auto-encoders, Bayesian (deep) neural networks, sequence models (LSTM/attention/transformers) and Gaussian processes. Your work would include familiarising yourself with one of these projects based on your preference, contribute to developing novel deep learning methods, and apply them to exciting real-world data from our national or international collaborators. Applicants are expected to have good knowledge of machine/deep learning, statistics, programming, and interest in bioinformatics and biomedicine. Research work can be continued after the summer. For more information, see (http://research.cs.aalto.fi/csb/publications) and relevant recent work e.g. (https://arxiv.org/abs/2006.09763 http://arxiv.org/abs/1909.01614 https://arxiv.org/abs/1912.03549  https://www.biorxiv.org/content/10.1101/542332v2) or contact Harri Lähdesmäki

____
 

 

 

39. Infinitely deep learning with differential equations 

Professor/Academic contact person for further information on topic: Prof. Harri Lähdesmäki
Contact info: [email protected]

We are looking for motivated summer interns to work on a new paradigm of deep learning, so-called infinitely deep learning. Infinitely deep learning methods replace the traditional deep learning methods (implemented with a limited/finite number of "layers") with continuous-time differential equations which are parameterized by deep neural networks, thus implementing deep learning methods that are "infinitely" deep. Popular examples include e.g. neural ODEs and continuous-time normalizing flows. The differential equation models can be parameterized by neural networks or (deep) Gaussian processes. These methods can be applied to standard classification and regression tasks but they can also be used to learn arbitrary continuous-time dynamics from data without any prior knowledge. Our current work in this topic includes e.g. (i) developing efficient Bayesian methods for robust learning of infinitely deep models from data, (ii) developing infinitely deep models to learn unknown/arbitrary dynamics of high-dimensional systems (e.g. in robotics, biology, physics or video applications) using a low-dimensional latent space representation, and (iii) developing infinitely deep models for reinforcement learning. The goal of the summer internship is to contribute to development of one of these models (based on your preference) and to implement these methods in e.g. Pytorch. The project requires good knowledge of machine/deep learning, mathematics, statistics, and programming. Research work can be continued after the summer.

For more information, see (http://research.cs.aalto.fi/csb/publications)  and relevant recent work

  1. e.g. (http://proceedings.mlr.press/v80/heinonen18a.html http://proceedings.mlr.press/v89/hegde19a.html
  2. https://papers.nips.cc/paper/9497-ode2vae-deep-generative-second-order-odes-with-bayesian-neural-networks
  3. https://arxiv.org/abs/2006.08956
  4.  https://arxiv.org/abs/2011.01226) or contact Harri Lähdesmäki

____

40. Computational techniques in the study of distributed algorithms

Professor/Academic contact person for further information on topic: Prof. Jukka Suomela
Contact info: [email protected]

Prerequisites: Good programming skills and familiarity with algorithms and theoretical computer science. Completing the course CS-E4510 Distributed Algorithms will be helpful but not necessary.

____

 

41. Infinitely deep learning with differential equations

Professor/Academic contact person for further information on topic: Prof. Harri Lähdesmäki
Contact info: [email protected]

We are looking for motivated summer interns to work on a new paradigm of deep learning, so-called infinitely deep learning. Infinitely deep learning methods replace the traditional deep learning methods (implemented with a limited/finite number of ""layers"") with continuous-time differential equations which are parameterized by deep neural networks, thus implementing deep learning methods that are ""infinitely"" deep. Popular examples include e.g. neural ODEs and continuous-time normalizing flows. The differential equation models can be parameterized by neural networks or (deep) Gaussian processes. These methods can be applied to standard classification and regression tasks but they can also be used to learn arbitrary continuous-time dynamics from data without any prior knowledge. Our current work in this topic includes e.g. (i) developing efficient Bayesian methods for robust learning of infinitely deep models from data, (ii) developing infinitely deep models to learn unknown/arbitrary dynamics of high-dimensional systems (e.g. in robotics, biology, physics or video applications) using a low-dimensional latent space representation, and (iii) developing infinitely deep models for reinforcement learning. The goal of the summer internship is to contribute to development of one of these models (based on your preference) and to implement these methods in e.g. Pytorch. The project requires good knowledge of machine/deep learning, mathematics, statistics, and programming. Research work can be continued after the summer. For more information, see

http://research.cs.aalto.fi/csb/publications

and relevant recent work e.g.

  1. http://proceedings.mlr.press/v80/heinonen18a.html, http://proceedings.mlr.press/v89/hegde19a.html,
  2. https://papers.nips.cc/paper/9497-ode2vae-deep-generative-second-order-odes-with-bayesian-neural-networks,
  3. https://arxiv.org/abs/2006.08956, https://arxiv.org/abs/2011.01226 or contact Harri Lähdesmäki

___

 

 

42. Learning resources for Programming Studios A and 2

Professor/Academic contact person for further information on topic: Prof. Lauri Malmi
Contact info: [email protected]

In this task, you would design and implement new programming exercises, including their model solutions and unit tests, as well as example programs for these two courses. The work requires good Scala programming skills and interest in developing good learning resources.

___

 

 

43. Investigating Fake News and its Spread in Social Networks

Professor/Academic contact person for further information on topic: Dr. Barbara Keller
Contact info: [email protected]

The world is full of information and hidden processes. We strive to understand the world better every day and in order to do so, we try to structure the available information. A crucial step towards such structuring is to simplify the information so that the underlying patterns become visible. Networks are an elegant and ubiquitous way to structure and represent information. Furthermore, they offer an intuitive way to illustrate data.

In this project we study large data sets with the means of networks. We investigate fake news and its spread and / or minority groups and their network structure. We focus, but are not limited to data from Facebook, Instagram and co-author networks.

The applicant is assumed to have solid programming skills (preferably in python), a good command of English and is interested to explore data and networks. No prior knowledge in graph theory, or about graph plotting software is required, although it might be helpful.

___

 

44. Developing the vHelix DNA nanostructure design platform

Professor/Academic contact person for further information on topic: Prof. Pekka Orponen
Contact info: [email protected]

The area of DNA nanotechnology [1] employs DNA as generic building material for assembling nanoscale objects with dimensions in the order of 10-100 nanometres. Our group has been developing, in collaboration with a biochemistry team from Karolinska Institutet, a general-purpose design platform ""vHelix"" for producing in particular 3D wireframe designs folded from a single long DNA strand [2].

A new, user-friendly and extendible version of the vHelix platform was developed as a summer internship project in 2020 (not yet launched for public distribution). After the 2015 publication of the DNA strand-routing algorithm [3] implemented in the current vHelix version, several alternative methods have emerged, and the goal of the present project would be to implement some of these more recent algorithms as plugins to the new extendible vHelix version.

The project requires familiarity with basic algorithm design techniques, facility with combinatorial thinking, and good programming skills. Previous knowledge of biomolecules is not necessary, although it is an asset. The work is performed in the context of research project “Algorithmic designs for biomolecular nanostructures (ALBION)”, funded by the Academy of Finland. For further information, please see the research group webpage at http://research.cs.aalto.fi/nc/.

  1. https://en.wikipedia.org/wiki/DNA_nanotechnology
  2. http://vhelix.net/
  3. Benson et al., Nature 2015, https://doi.org/10.1038/nature14586

____

 

45. Efficient visualisation and simulation of DNA strand-displacement systems

Professor/Academic contact person for further information on topic: Prof. Pekka Orponen
Contact info: [email protected]

DNA strand-displacement (DSD) systems are a fundamental technique in the research area of ""dynamic DNA nanotechnology"" [1], whose general goal is to implement dynamical behaviours, including computation, in biomolecular systems. The basic biochemical DSD mechanism, which comprises the controlled binding and release of short segments of single-stranded DNA, is quite simple, yet rich enough to support the design of general-purpose computational devices [2].

Because the molecular-level details of DSD systems can get quite complicated and the eventual chemical interactions rather intricate, high-level design, simulation and visualisation tools to support the design of complex DSD systems are in demand. Our group recently developed for this purpose a flexible rule-based modelling package RuleDSD, which at the simulation level integrates to the widely-used BioNetGen framework [3]. A graphical user interface to the RuleDSD package, including a simulated annealing -based strand network visualisation method, was developed as a summer internship project in 2020.

The first goal of the present project is to improve the efficiency and reliability of the RuleDSD strand network visualisation tool by replacing its general-purpose simulated annealing engine by an algorithmic method designed specifically for this task. A second goal is to address the methodological and efficiency challenges that arise in the current RuleDSD version when dealing with large, or even potentially infinite, strand networks.

The project requires familiarity with basic algorithm design techniques, facility with combinatorial thinking, and good programming skills. Previous knowledge of biomolecules is not necessary, although it is an asset. The work is performed in the context of research project “Algorithmic designs for biomolecular nanostructures (ALBION)”, funded by the Academy of Finland. For further information, please see the research group webpage at http://research.cs.aalto.fi/nc/.

  1. https://en.wikipedia.org/wiki/DNA_nanotechnology
  2. https://www.microsoft.com/en-us/research/project/programming-dna-circuits/
  3. Gautam et al., BIOSTEC 2020, https://doi.org/10.5220/0008979101580167

____

 

46. Pattern recognition from MHD simulations using Deep Learning

Professor/Academic contact person for further information on topic: Prof. Maarit Käpylä
Contact info: [email protected], [email protected]

The Pencil Code (PC), is a simulation suite for solving partial differential equations for magnetohydrodynamics, used for example when modelling the magnetic activity of the Sun that drives space weather. Ever increasing mesh sizes are used by the scientists to reach more turbulent regimes, corresponding to the conditions in the modelled objects themselves. Approaching the Exa-scale era, scientists performing such simulations are facing serious challenges, as it is not evident even how to store system states for re-starting, not to mention  auxiliary  data  for  more  detailed  data  analysis.  In  this  tricky  future prospect,  machine  learning (ML) can  be  of  great  help. The  long-term  aim  of  the project is to develop an online or offline structure-detector assistant for the PC.

The task of the summer intern includes: To continue developing an existing code based on the FasterRCNN object detection model. The code also includes a data augmentation pipeline, which is necessary for increasing the training data size and diversity; Generating training data for the neural network using idealised PC setups; use the generated training data for deep learning network; apply the trained network to detect the predefined structures and track their evolution in time from the real simulation data.

Prerequisites: Interest in applying ML in the domain of computational fluid dynamics is a strong bonus.  Basic knowledge on ML is required, and being familiar with toolboxes like PyTorch or Tensorflow is an extra benefit.  Fluency in a high-level programming language (e.g.  Python or C++) is required.

_____

 

47. 3D visualization of large-scale data from multi-physics simulations

Professor/Academic contact person for further information on topic: Prof. Maarit Käpylä
Contact info: [email protected], [email protected]

Large-scale simulations of, for example, magnetized fluids in stellar interiors produce huge amounts of three dimensional data, where each system state can comprise hundreds of Gigabytes or even Terabytes. Visualizing such data is challenging, and special tools are required. In this internship project, we propose to advance our existing Python framework, with which we are able to generate images by combining two-dimensional slices with 3D projections (https://owncloud.gwdg.de/index.php/s/iAq7VQ2Rb71Xfau#/files_mediaviewer/).

The task of the student intern includes further development of these tools, with or without parallel processing to handle multiple snapshots of large datasets for animation. We also require tools for 3D manipulation of the 3D arrays as a stand alone serial application utilizing PyVista, Paraview or other suitable software. For the large datasets we generate from our cutting edge simulations we shall ultimately require these tools to be operable on parallel computing platforms.

Prerequisites: Good knowledge in Python, and managing Jupyter notebooks. Some knowledge of supercomputing environments is a bonus.

____

 

48. AI-assisted Design

Professor/Academic contact person for further information on topic: Prof. Samuel Kaski
Contact info: [email protected], [email protected]

Historically the favored approach in machine learning has been to solve tasks be replacing humans entirely by ML systems. However both humans and ML systems have their own strengths. In recognition of this there has recently been increasing interest in creating hybrid Human-AI teams which aim to combine the strengths of both agents. In AI-assisted design we try to build such a team for solving design tasks, that is for optimizing a design according to an objective function known only by the human. In this internship you will work with us on one of the many aspects of this problem ranging from user modeling to planning and reinforcement learning. Requirements: solid math background, basic knowledge of RL and programming. Knowledge of cognitive science considered a plus.

Additional keywords: multi-agent learning, reinforcement learning, planning
Links: https://research.cs.aalto.fi/pml/

____

 

49. Byzantine fault tolerance with rich fault models

Professor/Academic contact person for further information on topic: Dr. Lachlan Gunn
Contact info: [email protected]

Byzantine fault tolerance (BFT) has seen a resurgence due to the popularity of permissioned blockchains.  Unlike with Bitcoin-style proof-of-work-based consensus, BFT provides immediate confirmation of requests/transactions.  Existing BFT protocols can generally tolerate a third of participants being faulty, unlike Bitcoin, which can tolerate attackers controlling up to a third of hash rate.
We are working to build a BFT system that can tolerate a richer variety of failure modes, for example:

  • Up to f nodes are malicious (this is the classical BFT)
  • Nodes with CPU power at most R are malicious (this is like Bitcoin)
  • All nodes running software X are malicious (e.g. a zero-day vulnerability is found in some piece of software)
  • All nodes owned by company X become malicious (e.g. someone steals administrator credentials)
  • Several of the above with different thresholds

In this project, we will develop systems that can tolerate more "real-world" types of fault like these, and gain experience in the development of distributed systems.
Requirements:

  • Basic network programming experience

Nice to have:

  • Familiarity with C++ and/or Rust
  • Theoretical distributed systems knowledge

____

 

50. Attack-tolerant execution environments

Professor/Academic contact person for further information on topic: Dr. Lachlan Gunn
Contact info: [email protected]

Modern processors are complex and incorporate several mechanisms like caching, speculation, and out-of-order execution to improve performance. Several recent attacks, like the well-known Spectre, Meltdown, and Foreshadow attacks, exploit this complexity to compromise integrity and confidentiality of computation.  We want to explore hardware and software innovations that can retain application data confidentiality even in the presence of such strong adversaries.  This will involve modifying existing open-source processor designs to develop new defences against runtime and side-channel attacks, and identifying and developing new ways of incorporating these features into real-world software.
Requirements:

  • C programming experience

Nice to have:

  • Basic cryptographic knowledge (hash functions, digital signatures, etc.)
  • Familiarity with computer architecture
  • FPGA development experience

____

 

51. Making high-integrity security policies usable in the real world

Professor/Academic contact person for further information on topic: Dr. Lachlan Gunn
Contact info: [email protected]

SELinux and other mandatory access control mechanisms have been used to secure operating systems, such as Android and normal Linux distributions. But with the great power of these mechanisms comes great complexity, and so far SELinux is used primarily by specialists as part of operating system development, and not by application developers. In this project we will remedy this by developing semi-automated and automated methods that analyse a system's access-control graph, using both data provided by application developers and heuristic application complexity measures, and provide useful data to the developer when deciding on their security boundaries. We will then incorporate these methods into a graphical tool that can be used by normal developers to design security policies that ensure the integrity of their applications.
Required skills:

  • Familiarity with Linux operating system security
  • Basic programming skills in your preferred language

Nice to have:

  • Familiarity with graph theory
  • GUI programming experience

____

 

52. Federated Learning: Adaptive Attacks and Defenses in Model Watermarking

Professor/Academic contact person for further information on topic: Prof. N Asokan
Contact info: [email protected]

Watermarking deep neural networks has become a well-known approach to prove ownership of machine learning models in case they are stolen [1,2,3]. Watermarking methods should be robust against post-processing methods that aim to remove the watermark from the model. Post processing methods can be applied by the adversary who stole the model after the model is deployed as a service.

Unlike traditional machine learning approaches, federated learning allows training machine learning models at the edge devices (referred as to clients or data owners) and then combines the results of all models into a single global model stored in a server [4]. Watermarking solutions can be integrated into federated learning models when the server is the model owner and clients are data owners [5]. However, unlike the post processing methods, adversaries as malicious clients can directly manipulate the model in the training phase to remove the effect of watermark from the global model. In this work, we are going to design an adaptive attacker that tries to remove the watermark in the training phase of the federated learning and propose new defense strategies against this type of attackers.

Note:

  1. You will be required to complete a programming pre-assignment as part of the interview process.
  2. This topic can be structured as either a short (3-4 month) internship or a thesis topic, depending on the background and skills of the selected student.

Required skills:

  • Basic understanding of both standard ML techniques as well as deep neural networks (You should at least take Machine learning: basic principles course from SCI department or some other similar course)
  • Good knowledge of mathematical methods and algorithmic skills
  • Strong programming skills in Python/Java/Scala/C/C++/Javascript (Python preferred as de facto language)
  • Sufficient skills to work and interact in English

Nice to have:

  • Familiarity with federated learning or statistics
  • Familiarity with deep learning frameworks (PyTorch, Tensorflow, Keras etc.)

References:

  1. Adi, Yossi, et al. 2018. "Turning your weakness into a strength: Watermarking deep neural networks by backdooring." 27th USENIX Security Symposium.
  2. Zhang, Jialong, et al. 2018. "Protecting intellectual property of deep neural networks with watermarking." Proceedings of the 2018 on Asia Conference on Computer and Communications Security.
  3. Darvish Rouhani et al. 2019. "DeepSigns: an end-to-end watermarking framework for ownership protection of deep neural networks." Proceedings of the Twenty-Fourth International Conference on Architectural Support for Programming Languages and Operating Systems.
  4. McMahan, Brendan, et al. "Communication-efficient learning of deep networks from decentralized data." Artificial Intelligence and Statistics. PMLR, 2017.
  5. Atli, Buse Gul, et al. 2020. "WAFFLE: Watermarking in Federated Learning." arXiv preprint arXiv:2008.07298

____

 

53. Latent Representations in Model Extraction Attacks

Professor/Academic contact person for further information on topic: Prof. N Asokan
Contact info: [email protected]

In recent years, the field of black-box model extraction has been growing in popularity [1-5]. During a black-box model extraction attack, adversary queries data to the victim model sitting behind an API and obtains the predictions. It then uses the data together with the predictions to reconstruct the model locally. Model exctraction attacks are a threat to the Model-as-a-Service business model that is becoming ubiquitous choice for ML offerings. Unfortunately, existing defense mechanisms are not sufficient and it is likely that model extraction will always be a threat[5]. However, despite the threat, research community does not fully understand why model extraction works and what are its current shortcoming and limitations.

In this work, we are going to explore the quality of latent representations learned during model extraction attacks, study the relationship between the victim and stolen models. We will investigate the impact of robustness-increasing techniques (e.g. adversarial training) on the effectiveness of model extraction and finally, formalise the field of model extraction attacks through the lense of transfer learning.

Note:

  1. There's going to be a programming pre-assignment.
  2. Literature review can be done as a special assignment.

Requirements:

  • MSc students in security, computer science, machine learning
  • Familiarity with both standard ML techniques as well as deep neural networks
  • Good math and algorithmic skills
  • Strong programming skills in Python/Java/Scala/C/C++/Javascript (Python preferred as de facto language)

Nice to have

  • Industry experience in software engineering or related
  • Research experience
  • Familiarity with adversarial machine learning

References:

  1. Nicolas Papernot, Patrick McDaniel, Ian Goodfellow, Somesh Jha, Z Berkay Celik, and Ananthram Swami. 2017. Practical black-box attacks against machine learning. In ACM Symposium on Information, Computer and Communications Security. ACM, 506–519.
  2. Florian Tramèr, Fan Zhang, Ari Juels, Michael K Reiter, and Thomas Ristenpart. 2016. Stealing machine learning models via prediction apis. In 25th USENIX Security Symposium. 601–618.

  3. Mika Juuti, Sebastian Szyller, Samuel Marchal, and N. Asokan. 2019. PRADA: Protecting against DNN Model Stealing Attacks. In IEEE European Symposium on Security & Privacy. IEEE, 1–16.

  4. Tribhuvanesh Orekondy, Bernt Schiele, and Mario Fritz. 2019. Knockoff Nets: Stealing Functionality of Black-Box Models. In CVPR. 4954–4963.

  5. Buse Atli, Sebastian Szyller, Mika Juuti, Samuel Marchal and N. Asokan 2019. Extraction of Complex DNN Models: Real Threat or Boogeyman? To appear in AAAI-20 Workshop on Engineering Dependable and Secure Machine Learning Systems

  6. Tero Karras, Samuli Laine, Timo Aila. A Style-Based Generator Architecture for Generative Adversarial Networks, in CVPR 2019

Department of Computer Science

See also other open positions in the department of computer science.

Department of Computer Science photo Matti Ahlgren

Open positions

Work with us!

Department of Computer Science
Computer Science building / Tietotekniikan talo / photo Aalto University, Matti Ahlgren

About the department

World-class research and education in modern computer science to foster future science, engineering and society. The work combines fundamental research with innovative applications.

Department of Computer Science
  • Published:
  • Updated:
Share
URL copied!