I am a research group lead at the Max Planck Institute for Intelligent Systems in Tübingen. My research focuses on the role of society in the study of computation, taking into account actions and reactions of individuals when analyzing and designing algorithmic systems. Prior to joining MPI I spent two years as an SNSF postdoctoral fellow at UC Berkeley hosted by Moritz Hardt. I have obtained my PhD from ETH Zurich where I was affiliated with the Data Analytics Laboratory and supervised by Thomas Hofmann. During my PhD I was employed at IBM Research Zurich where I contributed to the design and implementation of the IBM Snap ML library.
→ See my talk in the MIT OPTML++ seminar for an overview.
When informing consequential decisions, predictions have the potential to change the way the broader system behaves by triggering actions and reactions of individuals. Thereby they alter the data distribution the predictive model has been trained on -- a dynamic effect that traditional machine learning fails to account for. To better understand and address this phenomenon, we introduce the framework of performative prediction to supervised learning [ICML'20]. We analyze the dynamics of retraining strategies in this setup and address challenges faced in stochastic optimization when the deployment of a model triggers performative effects in the data distribution it is being trained on [NeurIPS'20]. When performative effects are strong we would wish to model and understand these effects in order to incorporate them into the very design of learning systems. Towards this ambitious goal we explore connections to microfoundations from macroeconomics theory and investigate how assumptions on individual behavior can be used to model and analyze performative effects in the context of strategic classification [ICML'21]. We study performative prediction through the lense of causal inference [NeurIPS'22], and we build on connection to online learning to design targeted experimentation and exploration strategies to collect data and find good models ex-post [ICML'22]. More broadly, performative prediction brings together ideas from machine learning, optimization, causality, and control theory to offer a formal framework to talk about the impact of algoirthms on society.
The extent to which predictions are performative is closely related to the economic concept of power -- the more powerful the firm making the prediction, the stronger the performative effects they induce. As such, power plays a suddle role in learning. It offers a lever to achieve low risk for the firm making the predictions by exerting influence on the population. We introduce the notion of performative power[NeurIPS'22a] to formally study power in prediction. We relate performative power to the economic study of competition in digital economies, analyze its role in optimization and propose an observational causal design to estimate performative power by exploiting discrete decisions of how firms display predictions, by building on properties of predictions more generally [NeurIPS'22b], and by building on connections to control theory to exploit repeated interactions between firms and individuals [CDS@NeurIPS'22]. In our most recent project we investigate algorithmic collective action as a means to counter power imbalances in digital economies. Most digital firms rely to some extent on data provided by individuals. This offers a lever to the population for gaining back some control over the system by strategically reporting data [Arxiv'23]. We study different learning settings and quantify the critical mass of individuals that need to be mobilized to achieve concrete goals. The critical mass is closely related to the cost of mobilizing a collective. Going further, I am interested in studying the effectiveness of concrete strategies and challenges of coordination, as well as connection to labor markets and collective action theory in political economy.
When training machine learning models in production, speed and efficiency are critical factors. Fast training times allow short development cycles, offer fast time-to-insight, and after all, save valuable resources. Our approach to achieving fast training is to enable the efficient use of modern hardware through novel algorithm design. In particular, we develop principled tools and methods for training machine learning models focusing on: compute parallelism [NeurIPS'19][ICML'20], hierarchical memory structures [HiPC'19][NeurIPS'17], accelerator units [FGCS'17] and interconnect bandwidth in distributed systems [ICML'18]. We demonstrated [NeurIPS'18] that such an approach can lead to several orders of magnitude reduction in training time compared to standard system-agnostic methods. The innovations of this research form the backbone of the IBM Snap ML library and help diverse companies improve speed, efficiency and scalability of their machine learning workloads.