Closing Remarks
Artificial Intelligence, Trustworthy AI, Counterfactual Explanations, Algorithmic Recourse
Around the beginning of this Ph.D. trajectory, my mother told me that apparently as a young kid I was fascinated with the 2001 movie, “A.I.”, by Steven Spielberg. According to her, at just 8 years of age, I must have been way more receptive to the potential impact of artificial intelligence than herself, evidently already paving the way for what would be my future in this field. Like most parents, I think my mother may have given me a little too much credit there, because all I remember from that movie is “the kid who played young Anakin Skywalker” (or at least that is who I thought it was1).
Still, after initially dreaming of becoming an actor myself only to eventually get a degree in economics and work in monetary policy for a while, somehow I ended up spending the past four years of my life researching the field that gave that Spielberg movie its name. So, how exactly does a former central banker end up pursuing a Ph.D. in Trustworthy AI?
My time at the Bank of England coincided with a number of organizational transformations that were geared towards embracing technologies and data sources, that in the eyes of most economists would have been considered unconventional at the time. The Bank’s Advanced Analytics (AA) division had been founded just three years before I entered the “Old Lady of Threadneedle Street”2 for the first time in the summer of 2017, then as an intern. While many of us at the Bank marvelled at the innovative research coming out of AA, its impact on policy decisions was probably fairly described as “tangential” (at least from my perspective as an analyst who regularly contributed briefing rounds of the Bank’s Monetary Policy Committee). Even though the terms “big data” and later “machine learning” would grab peoples’ attention during meetings, I sensed a certain reluctance amongst colleagues and superiors to substitute tried and trusted tools for new technologies that few of us understood very well.
At the time, I was naively optimistic that the necessary understanding could be swiftly acquired, and we would soon replace all of our ordinary least squares regressions with gradient-boosting trees and universal function approximators (a.k.a. artificial neural networks). Full of enthusiasm to “start building”, I went back to Barcelona School of Economics to study for a Master Degree in Data Science, and—as it turns out—to find out just how blatantly ignorant I had been about these promising new technologies. For the first time, I realized that the brave new world of machine learning clashed on fundamental levels with principles I had been taught in traditional econometrics: suddenly, we spent hours optimizing models for accuracy on Kaggle benchmark datasets with little to no attention to the underlying data itself. The assignment had completely changed: from understanding why things happen, to predicting what things happen. To be fair, we did cover explainability on the fringes even in these types of courses, but it still started to dawn on me that central banks would not, in fact, trust the Deep Vector Autoregressive Models we proposed in our master’s project to produce future inflation forecasts (Agustı́, Costa, and Altmeyer 2023). After all, public policymakers, and the people subjected to their decisions, do care about why things happen, not just what things happen. So, that is how I ended up pursuing this Ph.D. in Trustworthy AI, convinced—as I still am—that the trustworthiness of these models can and should be improved, even though it might not be easy.
Now at the end of this journey, I remain cautiously optimistic that we can continue to make progress towards trustworthy AI. Despite having witnessed practices and trends that concern me, I still believe it is possible to embrace and promote innovation and progress without adhering to premature paradigms like “moving fast and breaking things”. This will require patience, persistence and effort—virtues that I believe are being undervalued in many places and communities of today’s fast-paced world. Personally, I believe that trying to adhere to these virtues has played an important role in obtaining this Ph.D. degree, outweighed only by the importance of the many people involved in this journey.
Acknowledgements
A Ph.D. is often thought of as a somewhat lonely journey testing people’s ability to perform research in an independent and self-reliant manner. While this certainly applies in some sense and I have often enjoyed the freedom to work autonomously on topics of my choice, I have relied on countless people to get to this stage of the process. In this section, I want to acknowledge and thank some of these people.
First and foremost, I want to thank my partner, Daniëlle Sophie Kotter, who has worn many hats throughout this journey including ‘home office companion’, ‘therapist’, ‘interpreter’ and ‘travel buddy’, to name just a few. Having her by my side has made this all an overwhelmingly happy journey, notwithstanding the emotional and mental challenges that were occasionally and perhaps inevitably brought on by looming paper deadlines. Next to Dani, my friends and family have played a similar role in making me feel supported both in terms of my academic and professional aspirations and also outside that. Thank you all for always being there.
On the professional side, I want to begin by thanking Cynthia, who has been my daily supervisor. It has been said before that the relationship with supervisors can make or break Ph.D. students, and I would wholeheartedly agree with that. Cynthia has provided me with great academic freedom and autonomy, demonstrated faith in my work even when I had self-doubts, and—by setting such an admirable and impressive example herself—motivated me to approach this degree with the discipline and integrity it demands. Her scientific rigor, public outreach and genuine care for people, have always made me feel immensely privileged and proud to work under her guidance. Cynthia, thank you for everything you have done for me and everything you continue to do for the people around you! I am immensely grateful and continue to be amazed and inspired by everything you achieve. I am also sincerely grateful to Arie, who has consistently been supportive of all aspects of this project and frequently provided the necessary input to set individual projects on their right path. Arie has played an important role in ensuring that we present our work in the right manner and to the right audiences, which I am very grateful for. Especially during the early stages of this Ph.D., he has pushed me to aim higher than perhaps I would have, without inflicting unnecessary pressure. Thank you both, for always making me feel supported and appreciated, for motivating me through your interest in my work, and for creating many unique opportunities and connections.
Of course, I also want to express my gratitude to my many co-authors, colleagues and contributors. I have been asked once or twice why I always speak in the first-person plural when presenting work from my Ph.D. What by now has become force-of-habit, has always felt very natural to me: even though I feel a great sense of personal responsibility for the outcomes of this project, I believe that everyone who has been involved in one way or another also has a stake in it. I am very grateful to all of you and apologize to anyone I have missed: Jaehun Kim, Andrew Demetriou, Sandy Manolios, Marijn Roelvink, Imara van Dinten, Mojtaba Farmanbar, Elvan Kula, Leonhard Applis, Lorena Poenaru-Olaru, Floris den Hengst, Luís Cruz, Sara Salimzadeh, George Siachamis, Pasquale Caterino, Jorge Luiz Franco.
I also want to thank the many students at TU Delft I had the privilege to work with throughout this PhD. In particular, I want to highlight two exceptional students who I have not only had the privilege to supervise, but also directly collaborate with on research that is included in this thesis: Aleksander Buszydlik and Karol Dobiczek. Both have contributed as co-authors to Chapter 3 Endogenous Macrodynamics in Algorithmic Recourse after demonstrating remarkable skill and enthusiasm during their bachelor’s research projects. Two years later, Cynthia and I again had the pleasure of working with them on their master’s research projects, both of which have resulted in peer-reviewed publications. Finally, Aleksander also contributed substantially to Chapter 5 Counterfactual Training: Teaching Models Plausible and Actionable Explanations, once again demonstrating tireless enthusiasm and team spirit. Thank you both, I have really enjoyed working with you!
As I am finishing this thesis off for printing, I also want to express my gratitude to my Ph.D. committee and others who have provided valuable feedback and thoughts on thesis throughout: prof. dr. ir. Jan H. Kwakkel, prof. dr. Iman P.P. van Lelyveld, dr. Margaret Mitchell, prof. dr. Mykola Pechenizkiy, dr. Flavia Barsotti and dr. Jiahao Chen.
Last but definitely not least, I want to thank the open-source software community, most notably Julia and Quarto, for creating such a welcoming and supportive environment for researchers like myself. A huge proportion of my time spent on research has involved open-source software development, and both Julia and Quarto have made that time enjoyable. Julia and its community, I strongly believe, have made me a better developer and thereby enabled me to be a better researcher. Quarto has allowed me to prototype, produce and publish science in the way that I think it should be. Amongst other things, I have used Quarto to build this very thesis you are reading. I feel a deep sense of gratitude to both of these communities for making their products openly accessible.
As it turns out, the child actor playing the advanced robotic boy in “A.I.” is Haley Joel Osment, who auditioned to play young Anakin but did not actually get the role. Young Anakin was in fact played by Jake Lloyd. The more you know …↩︎
The Bank’s nickname dates back to cartoon published in 1797: https://www.bankofengland.co.uk/explainers/who-is-the-old-lady-of-threadneedle-street.↩︎