University of Tübingen

Middle East Technical University

EPFL, NeuroAI Lab | Research Intern

• Developing causal intervention methods for controlling internal representations in vision-language models for computational neuroscience.
• Linking model representations and interventions to human neural responses, with the goal of identifying shared and divergent representations between artificial and biological systems.

University of Tübingen, CVML Group | Research Assistant

• Developing JEPA-based multimodal alignment methods for self-supervised audio-visual representation learning.

METU ImageLab & Cambridge AFAR | Research Assistant

• Designed and implemented FairSSL, a self-supervised loss for modality-agnostic, fair multimodal representation learning tasks for neurological datasets (Joint work with Cambridge Affective Intelligence and Robotics Lab).
• Worked on model de-biasing strategies at the intersection of deep learning and kernel methods for fairness.

METU | Undergraduate Researcher

• Worked on causal agent-based models of decision theory in a medical shared decision-making setup.
• Modeled the conversation strategies employed by patients and doctors as a language game, consisting of various states, utilities, costs, and rewards.

METU | Undergraduate Researcher

• Applied Gaussian mixture models to inter-subject correlation features from fMRI to identify shared functional response regions.
• Modeled fMRI-derived functional connectivity using dynamic Bayesian networks to compare connectivity patterns between Alzheimer's disease and control groups.

GenVariableElimination.jl [Repo]

Extended the experimental variable elimination library of Gen.jl probabilistic programming system (GenVariableElimination.jl) with a JAX einsum backend for scalable factor-graph contraction on GPUs, preserving modeling interface and semantics across large models with over 10x speedup on specific tests.

Bayesian Normalization Layers Reproduction [Repo]

Implemented and reproduced the results of Bayesian Normalization Layers (BNL), CVPR 2024, contributing to the field as no implementation was available for this paper.