Bayesian analysis is a type of statistical methods. (Det store norske leksikon)
Østfold University College established a statistics tutoring service in collaboration with NMBU in August 2022. Scientific employee and fellows/PhD students at HiØ can send requests for statistical guidance. Aliaksandr Hubin is an associate professor in statistics. He can answer requests and provide statistical support and service.
Get to know Aliaksandr Hubin and his field of research in this interview:
My name is Aliaksandr Hubin, and I’m originally from Minsk, where I obtained a graduate degree in Applied Mathematics and Computer Science with a specialization in Econometrics and Statistics at Belarusian State University (BSU), which I finished with distinction.
It was a good school with inspiring peers around—many were medalists of the international mathematics (IMO) and informatics Olympiads (IOI), and our faculty's student teams regularly placed in the top 10 worldwide in the international collegiate programming contest (ICPC).
In August 2012, I moved to Norway, where I completed another Master of Science degree, again with distinction, followed by a PhD at the Department of Mathematics at the University of Oslo. I then worked as a research scientist at the Norwegian Computing Center, and as a postdoc at the University of Oslo.
Since August 2022, I have been employed as an associate professor in statistics (50%) at Østfold University College.
What inspired you to pursue a career in research, and how did you become interested in your specific field?
I have always been passionate about mathematics and its applications to real-world problems. During my studies in applied mathematics and computer science, I developed an early interest in mathematical modeling, particularly in probabilistic models for finance and combinatorial optimization for planning, routing, and control.
In my graduate thesis at BSU, I developed a Bayesian statistical model for asset pricing and portfolio management that took into account investors’ views.
Later, during my master’s degree in Norway, I worked on a problem simulating schedules for vessel fleets using discrete event simulation, considering both hard service rules and stochastic weather processes with potential online rerouting schemes via combinatorial optimization.
These projects combined various tools from applied mathematics, like statistics, simulation, and optimization, and involved extensive scientific programming.
For my PhD, I wanted to work on something more fundamental and general, which led me to focus on Bayesian statistics, specifically model and algorithmic developments for very general Bayesian non-linear models.
These included symbolic regressions, logic regressions, and neural networks as special cases, which led me into the subfield of Bayesian statistics that we now call probabilistic AI.
Could you describe your current research focus and the specific areas within AI that you are working on?
My research primarily focuses on Bayesian statistics and machine learning. Within what one calls probabilistic AI, I work on Bayesian model averaging for sparse variational neural networks and other sparse machine learning models in general, with a focus on sustainable, transparent, and uncertainty-aware solutions.
In other words, I develop methods that allow to create lightweight models that can indicate uncertainty levels in decision-making and are explainable by design, both for the population as a whole and for specific predictions.
I apply these methods to both tabular and image data. Recently, I’ve also worked on probabilistic state-space methods for natural language processing with applications to named entity recognition. Developing scalable computational solutions is fundamentally important in my work, so I focus heavily on creating efficient algorithms to solve these problems.
What motivated you to choose this particular research topic?
When I started my PhD, there were already informal discussions among researchers about the need for transparency and uncertainty handling in neural networks and other black box statistical predictive methods. These non-technical talks and discussions inspired me to pursue this topic further, and the issues remain relevant today.
The challenge of uncertainty in AI and the need for interpretable, reliable models naturally led me to Bayesian methods, which offer a flexible framework for handling model, parameter, and predictive uncertainty. This is especially useful in critical applications like healthcare or finance, where traditional AI models often are met with skepticism from professionals.
How long have you been working on this research, and what are some of the major milestones you have achieved so far?
I’ve been working in this field for roughly 10 years, with my interest and inspiration driving my progress.
Along the way, I completed my PhD and published numerous papers, including several in top-tier venues such as journals like the Journal of Artificial Intelligence Research, International Journal of Approximate Reasoning, Transactions on Machine Learning Research, and Bayesian Analysis as well as in the proceedings of major conferences like the international conference on machine learning (ICML) or the annual meetings of association for computational linguistics (ACL).
In Bayesian Analysis, my paper became the discussion article of the issue, with experts worldwide contributing discussions followed by a rejoinder from my team and a webinar. It was without doubt a proud personal achievement.
On the software side, I’ve developed two open-source R packages (FBMS and EMJMCMC) published on CRAN and contributed to the development of the Spacy library skweak in Python. The latter became quite popular in the industry as I have heard.
I also make all my papers' source code available on GitHub for reproducibility and for others to use my methods.
In addition to research, I serve the community by reviewing for reputable statistical journals and machine learning conferences. This year, I was recognized as one of the top reviewers at the UAI 24 conference, which was a pleasant recognition.
Lastly, as the chair of the Oslo section of the Norwegian Statistical Association, I organize many social and scientific events for the community, including last year’s European young statistician’s meeting and this year's national statistical meeting in Tønsberg with almost 100 participants.
Why do you believe your research is important for society?
My work enhances the sparsity, reliability, and interpretability of AI systems, which are becoming crucial in fields like healthcare, finance, and AI for science. These aspects should have been addressed before the AI boom driven by huge transformer architectures, which often lack these important features.
By addressing uncertainty and focusing on transparency, my research helps improve decision-making processes, making AI systems more trustworthy. Additionally, enforcing sparsity reduces the memory and computational demands of these models, making them greener and more sustainable.
Can you explain how your work can potentially impact or benefit everyday life?
Explainable-by-design AI systems that handle uncertainty effectively, like those developed in my research, can improve diagnostic tools in healthcare by offering transparent probabilistic assessments that may help doctors make better decisions. In finance, such models can provide more accurate risk predictions, improving investment, policy decisions, or credit approval.
Proper uncertainty handling in areas like robotics and self-driving cars can also reduce life-threatening risks.
While my work is often not directly applied per se, I do not find myself in the position of a pure mathematician, who knows his work will unlikely be applied and sometimes even praises this (as a joke, of course), the potential applications of my models can benefit everyday life if adopted by industry stakeholders. By all means, I hope they are used more in science, industry and society.
What are some of the challenges you face in your research, and how do you overcome them?
One significant challenge is the computational complexity of Bayesian models, especially when applied to large datasets. To address this, I focus on developing and applying scalable methods like scalable stochastic variational inference and subsampling Monte Carlo Markov chains (MCMC). I also stay updated with the latest techniques by attending and presenting at relevant conferences.
Another challenge is the state-of-the-art-driven publication bias in top-tier conferences, where small non-significant performance gains often overshadow fair experimental designs and foundational research.
I believe this is a worrying trend, and I hope the community will work towards addressing it. I as a reviewer try to do my best to improve this at least on a local scale.
In what ways do you think AI can transform different sectors of society, and what potential ethical considerations should be addressed?
AI may revolutionize sectors like healthcare, agriculture, finance, and even scientific discoveries by automating processes, enhancing decision-making, and generating new insights. For healthcare, specifically, personalized AI-enhanced treatments seem extremely promising.
However, considerations such as data fairness, statistical guarantees of algorithmic convergence and model generalization, and transparency for decision-making under uncertainty must be addressed to ensure that AI systems are trustworthy statistical models that align with societal values.
I strongly believe that those issues should have been properly addressed before allowing us to move to the scale of modern AI models, which are unsatisfactory and wild in many senses.
Thoughtful regulation was (and is still in many countries) lacking, which allowed this to happen. And thoughtful regulation is also essential for the future in my view. After all, it is important to remember that powerful AI tools are still used by humans whose intentions are not always good, sadly.
To name a few, fake news, political propaganda, and mass manipulation of public opinion are now often done using AI.
What are your future research plans, and what do you hope to achieve in the next few years?
I plan to continue advancing my work on Bayesian machine learning and deep learning, particularly by developing mode-jumping Monte Carlo Markov chain samplers for neural networks and creating prior constructions to break posterior symmetries and further improve interpretability.
Mentoring will remain a central part of my work, guiding master’s and PhD students through these cutting-edge areas of research. I hope to establish my own mini-lab and attract bright students and researchers at some point.
I also aim to develop and teach a course on Bayesian machine learning and deep learning methods, which would be novel for Norway. Being part of an international team writing the first textbook on Bayesian deep learning, I would integrate this knowledge into the course.
Ultimately, I’d like to become a full professor with a full-time position. But in the short term, let us say I just plan to focus on high-quality and reasonable quantity research and student mentoring, while maintaining a good work-life balance.
How do you stay motivated and inspired in your research journey?
Natural curiosity has always been a source of motivation for me. This was since childhood when I would spend days experimenting in designed competitions for toy cars, recording the results and establishing the winners.
Also, mentoring students and watching them grow as researchers is deeply fulfilling, especially when they contribute to cool projects. The evolving challenges in AI and Bayesian methods keep me intellectually stimulated, and networking with colleagues working on similar problems is also inspiring.
Do you have any advice for young researchers or students who are interested in pursuing a career in AI and research?
I feel still young in my early thirties. For younger researchers, however, my advice is to build a strong foundation in mathematics, statistics, and programming.
Avoid following the hype surrounding popular topics like large language models or diffusion until your fundamentals are strong. AI is an interdisciplinary field, so a solid technical background will allow you to tackle complex problems with confidence.
Mentorship is also invaluable — I’ve seen students thrive by seeking and getting advice and guidance. I also encourage young researchers to dare to submit their work to top venues, get visible, and actively participate in conferences to build networks.
Your peers will be the future leaders in the field, so investing in relationships now will be beneficial later.
And simply be nice, it is important.
Read more about Aliaksandr Hubin and his research
Østfold university colleges research investment area “The digital society”