How did your adventure with science begin?
My journey with science began at quite an early age. I have always been interested in how everyday things work, taking apart things like electronic toys or kitchen mixers, to see their inner workings. As I got older my interests evolved to programming and electronics, making my own circuit boards, and coding plugins for games when I was in secondary school. I decided to study computer science so that I could develop a more in-depth and grounded understanding of how computers work. During a placement year at university, I worked with High Performance Computing (HPC) and server infrastructure, which gave me a deeper appreciation of the scale that is needed to power the modern world, and now the Artificial Intelligence (AI) revolution. I took a module at university which was an introduction to AI explaining how neural networks are trained from a single neuron to exceptionally large networks. Then I knew I wanted to pursue it further, which is why I applied to do a PhD in AI.
What factors influenced your decision to go abroad? What were the biggest challenges related to this decision?
I was always interested in living abroad, and experiencing new cultures and perspectives, so I applied to study in the UK. Since then, I had the opportunity to make new lifelong friendships and work with great academic supervisors. I was lucky to have an undergraduate project supervisor who encouraged me to do a more research oriented dissertation, which focused on cybersecurity applications of quantum computers. This is what made me strongly consider applying to a PhD programme, since I enjoyed that research experience and the research culture in the UK. I am happy with my decision as I now have a knowledgeable and supportive supervisory team, from whom I have learned a lot!
In terms of the challenges – I think in general, moving away from home is challenging. I was further from my support network of friends and family, which was especially hard when I first arrived, since I did not know anyone. While I tried to stay connected with friends who were back in Poland, I also had to balance that with forming new relationships with people I met. This was challenging as – and I think many new expats can relate to this – making your second language your everyday language can be scary at first. It was also challenging to familiarise myself with the diversity of UK accents while I was still getting used to using English as my new primary language, but this got easier over time and with experience. There were also practical challenges, such as figuring out the housing situation (guarantors?!) and local norms that took a while to adjust to, such as driving on the left. Now I must remind myself that jaywalking is illegal whenever I go back to Poland! Knowing that I have been able to overcome these challenges gives me more confidence in my ability to adapt to new environments, and I now feel more comfortable considering moving to other countries. My advice for anyone considering such a move is to be open to change and embrace the new culture and experience as much as possible!
What benefits does work in an international scientific environment bring you compared to working in Poland?
I thoroughly enjoy having the different perspectives that working in an international environment provides. I enjoy the mix of cultures and opinions which give me insights I would not have thought of on my own, enriching both me as a person, as well as my scientific research. Unfortunately, I have not had the opportunity to work in Poland, so I am not able to make that comparison yet, though I do hope to explore more collaborations with researchers soon, whether Polish or otherwise!
Do you maintain contact with Polish scientific communities?
Since I have both gone to university and am now doing a PhD in the UK, most of my scientific network is based here. I do, however, know of exciting work from the Polish scientific communities, including in my field. I have read some interesting research in emergent communication from Polish researchers, such as the analysis of the impact of noise on emergent communication by Kucinski et al. [1], or new ways of studying the emergent languages by Korbak et al. [2]. I hope that in the future I will be able to collaborate more with the Polish scientific community!
What are you currently working on and what is the main subject of your scientific research?
I work in AI, and more specifically in a field called emergent communication. Quite literally, I study the development of communication between autonomous AI agents. These agents are like players in a virtual world (such as a computer game) who are working together on various tasks. They learn about the task through reinforcement learning, where they receive a reward if they perform well in a task, but no reward or a penalty if they perform poorly. Although these agents are autonomous and can act independently of one another, they are working to reach a common goal and thus need to communicate. However, they are not told what to say or how to convey the information, so they need to develop their own language from scratch. This involves the agents learning what information is important to tell the others and the most efficient way to convey it. The resulting language is therefore usually specific to the environment that they were in. An advantage of allowing the agents to develop the language among themselves is that it is usually faster than instructing the agents on how to communicate and the resulting language is more efficient. This means that less programmer input is needed, increasing their work efficiency, and increasing the AI energy efficiency thanks to a language that is more efficient to transmit. My research specifically focuses on temporal references in emergent communication. Temporal references are words that would be able to describe relative relationships across time, such as “before” or “after”. Given the emergence of such words, agents can describe the events that they observe more precisely, allowing for better experience sharing between them.
What are the latest achievements in your field of research that particularly interest you?
In AI generally, I think that the pace of progress being made in generative AI models, such as ChatGPT or DALLE-2, is fascinating. They allow people to be more productive and improve our everyday lives. I, for example, often use ChatGPT for coding help or even brainstorming research ideas! The progress along the path to AGI (Artificial General Intelligence, an AI that could match or exceed human capabilities in many tasks) is also remarkable, with a lot of highly accomplished AI researchers working on this problem. While there is quite a way to go (especially when seeing how many mistakes ChatGPT can make), I think I may be able to see it in my lifetime.
In emergent communication, I think the path towards more interpretability is critical. A lot of AI systems are treated like black boxes, and we have very little idea how they make decisions. Being able to understand what agents communicate about would give us at least some insights into what they find most important. This in turn would make AIs safer, by making sure that their goals are aligned with ours. It would also make them more trustworthy and accountable, since we would know better what tasks they are working on and why. For example, I have read work from Ueda et al. [3], where they use tools developed for the study of human language or linguistics. They decompose the AI language into its constituents and analyse how they can be combined to create meaning. While there is still a way to go, before we can fully understand the emergent language, it is a very promising avenue of research. It is, however, very exciting to be able to apply insights gained from natural language to study something as “alien” as a language developed by an AI.
What is your most important scientific achievement or discovery? Why is it important?
So far, my biggest achievement is having shown the ways that agents can develop temporal references. We recently discovered that making a minor change way that AI agents work, is enough for them to learn to communicate about temporal relationships. This means that allowing agents to communicate about relationships across time is easy to transfer to other AIs. It could bring higher efficiency to the way agents communicate – for example, an autonomous vehicle would not need to describe the same junction every time it drives through it, it could instead say that this is the same junction as yesterday.
Which scientific problems in your discipline are you most looking forward to being solved and why?
In terms of general AI, I am very curious to see when more powerful AIs will be capable of exceeding human capabilities. I think that AGI could significantly change how we live our lives, by accelerating all our scientific progress, like drug discovery, understanding and treating complex diseases, or even helping solve climate change. There is ongoing debate about how to safely continue the developments in this field, because of concerns around making sure that an AI smarter than any human is still aligned with human values. And even pioneers of AI, such as Yann Le Cun or Yoshua Bengio, have not reached consensus on the best way forward. I would prefer to stay on the side of caution and more thoroughly evaluate the risks of AGI before such AIs are developed and deployed in the real world. But I am nonetheless excited about what the future will hold.
In emergent communication, I think that being able to fully decode emergent languages will be an important milestone in the field. These agents are artificial constructs and quite unlike humans, so it would be like being able to automatically decode and understand an alien language. This could provide insights on how linguistic concepts like grammar have emerged naturally in different human languages, as well as improve our understanding of how the AIs themselves work.
What are the biggest challenges you face in your scientific work?
The greatest challenge in my own work is writing up my research findings. I am fortunate to be working with supervisors who provide very helpful feedback on my drafts and have also received insightful guidance from other researchers through submitting my work to various academic conferences. However, no matter how many times I revise the work, there is always room for improvement!
What are the most important research questions you plan to address in the near future?
I am currently working on creating a way of automatic translation of emergent languages. As I have mentioned before, we do not really know how AIs make their decisions. Additionally, a lot of the methods used so far for understanding emergent languages rely on some knowledge of what they can see in their environments. I am hoping to find a way to significantly reduce or remove the reliance on this “insider” knowledge and be able to assign meanings to the agents’ language without human input.
Are there practical consequences or potential applications of the results of your scientific research? How do you see their impact on society or the economy?
I believe the wider impact of developments in AI will be far-reaching, even at its current stage. Even the current generation of Generative AI could affect the job market, by changing which jobs are more or less in demand, as well as helping most workers be more efficient, by reducing the need to perform repetitive tasks and increasing automation. We are already seeing the beginnings of this, with the widespread adoption of using AI to generate images, even by conventional media outlets, like Reuters using AI to generate video reports [4]. In longer term future impacts, I think that if AI does achieve human-level general intelligence (and it looks to be heading that way), societal structures overall could be affected. For example, a current concern is that our reliance on certain jobs could reduce or shift in focus if AGI is able to perform these roles more efficiently, which would remove the need for humans to be employed in them. Ironically, the jobs predicted to be automated by AI the fastest are the ones which are currently benefitting from the AI boom, such as programmers [5].
In terms of emergent communication specifically, I think the resulting outcome will be improved efficiency in communication between agents, and therefore reduced energy usage for any company or people using communicative AIs. This in turn could reduce their cost to both the user, and the planet in terms of carbon emissions. Hopefully, with enough progress in the field of interpretability, this communication will also be more understandable by us humans. This would give us the best combination, where naturally developed communication between agents is more efficient and adaptable than programmer-instructed communication, while also allowing us to understand the information communicated.
What advice would you give to young scientists at the beginning of their scientific careers?
The best advice I have received and would like to pass on is to follow your passion. A scientific career involves a lot of rejection – not just from job applications, but also from applications to publish your work in academic journals and conferences. Anyone wanting to pursue a scientific career needs to be persistent and resilient in the face of these challenges, but it is even more difficult to do that if you are not passionate about the work you are doing. In my experience, and from speaking to other academics, it is the enthusiasm and drive we have that motivates us to keep pursuing our research interests through our PhDs and beyond.
Bibliography:
[1] L. Kucinski, T. Korbak, P. Kolodziej, and P. Milos, ‘Catalytic Role Of Noise And Necessity Of Inductive Biases In The Emergence Of Compositional Communication’, NeurIPS 2021
[2] T. Korbak, J. Zubek, and J. Raczaszek-Leonardi, ‘Measuring non-trivial compositionality in emergent communication’, 4th Workshop on Emergent Communication, NeurIPS 2020
[3] R. Ueda, T. Ishii, and Y. Miyao, ‘On the Word Boundaries of Emergent Languages Based on Harris’s Articulation Scheme’, in ICLR 2023
[4] https://www.forbes.com/sites/simonchandler/2020/02/07/reuters-uses-ai-to-prototype-first-ever-automated-video-reports/ [5] https://www.forbes.com/sites/forbestechcouncil/2022/02/18/15-jobs-and-tasks-tech-experts-believe-will-be-automated-within-a-decade/?sh=5339d5a4778a
Fot. Unsplash
