Ata Atabek

The central objective of this PhD is to obtain rigorous training in neuroeconomics while developing principled machine-learning methods that address core empirical challenges and open new research directions. The research investigates decision-making under uncertainty across three interconnected levels of analysis: neural activity, individual behavior, and collective dynamics. The overarching aim is to integrate empirical decision neuroscience with modern machine learning to build models that are not only strongly predictive but, where possible, mechanistically informative. The project is structured as a four-part research program, with two projects currently underway and two additional projects now concretely defined.

The first project uses fMRI to characterize multisensory salience processing and develops neural-network models to capture its computational structure. The second project designs disentangled recurrent neural network architectures that infer latent cognitive factors from online decision-making behavior, replacing handcrafted cognitive models with unsupervised, data-driven representations. This architecture will later be used to analyze temporally unfolding social decision-making games, including repeated strategic interactions such as Rock-Paper-Scissors (RPS), allowing disentangled latent-state tracking in interactive settings. The third project develops a formal mathematical theory of disentangled recurrent architectures, with the goal of informing neuroscientists about principled conditions under which latent cognitive factors can be separated, interpreted, and generalized across tasks. This theory-first component provides the analytical foundation for the empirical modeling pipeline and establishes testable criteria for model identifiability and robustness. The fourth project advances the disentangled recurrent architecture to enable real-time disentanglement in increasingly complex tasks such as vision tasks, with the explicit goal of transferring these capabilities to salience-oriented paradigms. Taken together, the PhD program targets three complementary challenges in neuroeconomics: (i) mechanistically grounded hypothesis generation at neural and cognitive levels, (ii) mathematically principled disentanglement of latent cognitive structure, and (iii) scalable, real-time modeling of complex perceptual and social decision processes. By embedding machine learning directly into experimental design, inference, and post hoc computational analysis, the project addresses longstanding methodological limitations in neuroeconomics while opening multiple new avenues for empirically grounded, computationally principled research.