22 postdoc, research fellow and PhD student positions in artificial intelligence and machine... (# of pos: 20)

Updated: 2 months ago
Job Type: FullTime
Deadline: 05 Oct 2020

Finnish Center for Artificial Intelligence FCAI offers a possibility for 22 new researchers to join a unique research community with an attractive joint mission.

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 is built on the long track record of pioneering machine learning research in Helsinki with currently over 60 professors from Aalto University and University of Helsinki contributing to our research. Our lively AI research community  organizes frequent seminars with prominent speakers and offers high-quality collaboration opportunities with other leading research networks and companies (e.g. FCAI hosts ELLIS unit Helsinki and has a joint research center with NVIDIA ). Local and national computational services spearheaded by the future EuroHPC supercomputer LUMI (200+ Pflops/s) provide our researchers with access to excellent computing facilities.

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. This is advanced in seven Research Programs and in Highlights that make sure FCAI’s fundamental AI research is taken into real-life use. Read more about our research here .

We provide an exceptional opportunity for 22 new researchers to join the FCAI community, where a great number of experts from a range of fields work towards the common research mission. The new positions are designed to further strengthen this research from diverse angles and offer a possibility for across-field collaboration through joint supervision by FCAI professors.

Depending on the position, we are looking for mainly postdocs, but also excellent research fellow and PhD student candidates can be considered for some of the positions. The positions are negotiated on an individual basis and may include e.g. relocation bonus and independent travel budget. We welcome applicants with diverse backgrounds, and qualified female candidates are explicitly encouraged to apply.

The deadline for applications is October 5, 2020 (midnight UTC+02:00)

Topics with a special emphasis on AI across fields collaboration


TOPIC 1: ACTIVE LEARNING FOR DECISION-MAKING

While basic active learning is usually applied for acquiring labels of unlabeled samples for supervised learning, the principle is more general. We develop methods for deciding what to measure or what data to collect, for a decision-making task which is initially partially unknown and has to be learned as well. Prime examples are learning of models of decision makers, or reward functions. The methods are central in developing AI-assisted decision-making tools. We provide good opportunities of applying the tools in personalized healthcare and medicine.

  • Supervision: Samuel Kaski (Aalto University), Pekka Marttinen (Aalto University)

  • Keywords: Active learning, decision-making, sequential design of experiments

  • Most relevant FCAI research programs: Agile probabilistic AI; Autonomous AI, Simulator-based inference; Next generation data-efficient deep learning; Interactive AI;  Easy and privacy-preserving modeling tools

  • Level: Postdoctoral researcher or research fellow


TOPIC 2: AI FOR THERAPY OPTIMIZATION

Predictability and control of evolving populations is an emerging topic of high scientific interest and vast translational potential in applications such as vaccine design and cancer therapy [1,2]. This project will develop inference approaches to optimally control dynamical systems (say cell populations) in a (partially) model free setting, the killer application being to derive optimal control (therapy) protocols for minimal models of cancer or microbial cell populations, without prior knowledge of the evolution equations. We assume an access to a simulator that can be used to generate outcomes of the process under some control protocol that is decided by us. Then we need to find an iterative solution that tells how to update the sampling control before generating another set of realisations from the simulator. It will be critical that not many iteration rounds are needed, so Bayesian optimization/ ELFI will be used and developed further [3]. The simulator can be at a first instance computational but can also be thought as a real biological experiment which can be further pursued with experimental collaborators.  

  • Supervision: Ville Mustonen (University of Helsinki); Jukka Corander (University of Helsinki)

  • Keywords: Stochastic optimal control, likelihood-free Inference, bayesian inference

  • Most relevant FCAI research programs: Simulator-based inference, AI for applications in healthcare

  • Level: Postdoctoral researcher

[1] Lässig M, Mustonen V, Walczak AM (2017) Predicting evolution. Nat Ecol Evol 1(3):1–9.
[2]  Lässig M, Mustonen V, (2020) Eco-evolutionary control of pathogens. PNAS, (in press).
[3]  Lintusaari J, [6 authors], Vehtari A, Corander J, Kaski S  (2018) ELFI: Engine for likelihood-free inference. J Mach Learn Res 19:1–7. 


TOPIC 3: AI-ASSISTED DESIGN OF INTERVENTIONS

The problem of choosing interventions is an exciting combination of causal inference and sequential experimental design, needed for both experimental scientific research and choosing medical treatments, for instance. We additionally take a human expert in the loop, having prior knowledge which can be queried, and whose goal is to learn about the causal effects in the process.

  • Supervision: Samuel Kaski (Aalto University), Kai Puolamäki (University of Helsinki), Antti Oulasvirta (Aalto University)

  • Keywords: Causal inference, sequential experimental design, interactive machine learning

  • Most relevant FCAI research programs: Agile probabilistic AI; Autonomous AI; Easy and privacy-preserving modeling tools; Simulator-based inference; Interactive AI; Next generation data-efficient deep learning

  • Level: Postdoctoral researcher or research fellow


TOPIC 4: AI-ASSISTED MODELLING

Modelling is a combination of a design task of building the model, and the statistical task of fitting the model to data or more generally statistical inference. While probabilistic programming is progressing fast on the latter part, and AutoML helps when there is enough data, much less help exists for the design task. We formulate the task of offering design help as a broader modelling task, which includes the modeller in the loop. The solution of the broader task gives both the AI-assistance and a solution of the primary task.

  • Supervision: Samuel Kaski (Aalto University), Aki Vehtari (Aalto University), Arto Klami (University of Helsinki), Antti Oulasvirta (Aalto University)

  • Keywords: Sequential design of experiments, prior elicitation

  • Most relevant FCAI research programs: Agile probabilistic AI; Interactive AI;  Easy and privacy-preserving modeling tools; Simulator-based inference; Next generation data-efficient deep learning

  • Level: Postdoctoral researcher or research fellow


TOPIC 5: ATMOSPHERIC AI

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 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. The AI part of the project will be executed in collaboration with the relevant FCAI AI research programmes.

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

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

  • Most relevant FCAI research programs: Agile probabilistic AI; Simulator-based inference; Interactive AI; Intelligent urban environment

  • Level: Postdoctoral researcher or PhD student


TOPIC 6: COMPUTATIONAL COGNITIVE MODELS

To reason about human behavior, and to plan actions and interventions, a strong prior is needed. The goal of this project is to develop computational models of basic cognitive capabilities, such as attention, memory, or decision-making. We in particular focus on rational models (e.g., bounded rationality, computational rationality) that can approximate behavioral policies. These models can support interactive AI in many ways, including training of ML models  (e.g., offline reinforcement learning), planning problems (e.g., model-based reinforcement learning), and inference of human data (e.g., using ABC). 

  • Supervision: Antti Oulasvirta (Aalto University); other possible supervisors Samuel Kaski (Aalto University), Jukka Corander (University of Helsinki), Ville Kyrki (Aalto University)

  • Keywords: Computational cognitive models, cognitive simulations, computational rationality, interactive AI

  • Most relevant FCAI research programs: Interactive AI; Simulator-based inference; Next generation data-efficient deep learning; Easy and privacy-preserving modeling tools; AI-driven design of materials

  • Level: Postdoctoral researcher


TOPIC 7: DATA-EFFICIENT DEEP LEARNING FOR ROBUST RETINAL DISEASE DETECTION

The goal is to develop and apply data-efficient Deep Learning (DL) methods for robust detection and classification of retinal diseases: diabetic retinopathy, diabetic macular edema, age-related macular degeneration, and glaucoma [1]. The focus is to increase trust in and reliability of DL methods in the decision-critical medical domain and its applicability for small annotated datasets, as well as developing robust DL for improved model calibration in on-going collaboration with the Alan Turing Institute and Helsinki University Hospital (HUS). The data-efficiency is explored by semi- and self-supervised learning and in difficult cases by adding dimensions through additional patient or multimodal data. Our medical collaborators in the Central Finland Central Hospital, Folkhälsan, HUS Eye-clinic & Glaucoma Center share large and unique sets of annotated Retinal, Optical Coherence Tomography (OCT) and other medical images, with auxiliary clinical patient data (such as patient’s blood glucose and disease & treatment & medication history data) but also large quantities of unannotated image data. The utilization of additional data would not only amount for novel studies in the domain of medical image analysis, but when applied, it could benefit the performance of such algorithms and models. The aim is that the deep learning systems are to be designed with uncertainty quantification and auxiliary medical information in mind, which would enable the system to recommend certain diagnoses to the medical expert and also quantify the certainty of recommendation. Due to having access to a number of very large and medically unique datasets of various kind (HUS, Central Finland's Central  Hospital, Digifundus Ltd, Folkhälsan), one of our additional goals is to develop deep learning methods for generating synthetic differentially private dataset for the purpose of wide-scope AI-methodology research, medical doctor training, and digitalisation of health care. 

  • Supervision: Kimmo Kaski (Aalto University), Simo Särkkä (Aalto University), Antti Honkela (University of Helsinki), Pekka Marttinen (Aalto University), Samuel Kask i (Aalto University)

  • Keywords: Deep Learning, robust learning supervised, semi-supervised and self-supervised methods, model calibration, AI-assisted medical decision-making, synthetic differentially private data

  • Most relevant FCAI research programs: Next generation data-efficient deep learning; Applications of AI in healthcare; AI's for AI-assisted design, decision-making and modelling; Security and privacy

  • Level: Doctoral student or postdoctoral researcher

[1] Sahlsten J, Jaskari J, Kivinen J, Turunen L, Jaanio E, Hietala K, Kaski K (2019), Deep Learning Fundus Image Analysis for Diabetic Retinopathy and Macular Edema, Nature Scientific Reports, 9, 10750. 


TOPIC 8: DEEP LEARNING IN MATERIALS ANALYSIS

Many challenges in interpretation of materials data are obviously suited to AI-assisted techniques and great progress has been made recently in data-mining with materials’ simulations. A step beyond this, is the introduction of AI tools into the direct analysis of experimental materials


View or Apply

Similar Positions