Abdulkadir Gokce

I am a second-year Ph.D. student in Computer Science at the Swiss Federal Institute of Technology Lausanne (EPFL), where I am advised by Martin Schrimpf. My research focuses on building artificial models of perception.

Previously, I earned my master's degree at EPFL and my bachelor's in Electrical & Electronics Engineering and Mathematics at Bogazici University. I was fortunate to gain research experience at ETH Zurich, MIT, and Nanyang Technological University (NTU).

Email  /  Google Scholar  /  Github  /  Twitter  /  Bluesky

Research

My research investigates the extent to which current deep learning models can capture neural and behavioral signals, and aims to develop new modeling strategies that expand these capabilities. I am particularly focused on building scalable, multimodal models of human perception that integrate diverse neurocognitive datasets.

First-Authored Papers

* indicates equal contribution.

We introduced MIRAGE, a whole-brain fMRI encoding framework that combines a native multimodal foundation model with adaptive layer gating, achieving state-of-the-art prediction of brain responses to naturalistic movies.

We mapped the scaling laws of brain alignment across vision models and neural recordings, showing that bigger pretraining helps only up to a point, after which neural supervision and improved mappings become the main drivers of progress.

We systematically explored scaling laws for primate vision models and discovered that neural alignment stops improving beyond a certain scale, even though behavior keeps aligning better.

Inspired by human cognitive development, our BabyLLaMA model learns language and vision jointly through a self-synthesis loop, generating its own training data from unlabeled images.

Co-Authored Papers

We developed a model-guided framework that predicts and controls how microstimulation in primate visual cortex biases perceptual choices, showing strong correlation between model predictions and behavioral changes in macaque monkeys performing visual recognition tasks.

We found that LLMs capture aspects of the neural geometry of human creative thought, and that creativity-focused post-training preserves alignment with highly creative neural responses while reasoning-focused training shifts it away.

We find that contour integration, a core feature of human object recognition, emerges in models only at large scales and correlates with improved shape bias.

We found that the most brain-like video models capture human visual cortex through two core computations — object recognition and appearance-free motion recognition — which together organize visual processing across cortical streams.

EndoL2H enhances capsule endoscopy images up to 12x using a spatial attention-guided GAN, outperforming existing methods in perceptual quality and clinical relevance.