PhD Oral Exam - Sana Ahmadi, Computer Science

When studying for a doctoral degree (PhD), candidates submit a thesis that provides a critical review of the current state of knowledge of the thesis subject as well as the student’s own contributions to the subject. The distinguishing criterion of doctoral graduate research is a significant and original contribution to knowledge.

Once accepted, the candidate presents the thesis orally. This oral exam is open to the public.

Abstract

This thesis investigates techniques for optimizing brain encoding models, em- phasizing computational efficiency and the scalability of both data and mod- els within the framework of large-scale functional magnetic resonance imaging (fMRI) datasets. Brain encoding aims to predict neural responses to complex stimuli, such as video frames, by utilizing latent feature representations from artificial neural networks. The first study explores the acceleration of ridge re- gression, a widely used predictive model in brain encoding, particularly when applied to large fMRI datasets like the CNeuroMod Friends dataset. By imple- menting a novel batch-parallelization strategy using Dask, we achieved significant computational speedups of up to 33× with 8 compute nodes and 32 threads com- pared to a single-threaded scikit-learn.

The second study investigates how dataset size and model scaling affect brain en- coding performance using vision Transformers. To do so, the VideoGPT model was trained end-to-end to extract spatiotemporal features from the Shinobi video game dataset with varying sample sizes (10K, 100K, 1M, and 6M) and model size (number of training parameters). Ridge regression is then used to predict brain activity based on fMRI data and the extracted features from video games. Our re- sults show that larger datasets lead to significantly improved encoding accuracy, with the 6M-sample dataset producing the highest Pearson correlation coeffi- cients across subjects. Additionally, while increasing hidden layer dimensions in the transformer model greatly enhances performance, the number of attention heads appears to have a minimal effect. These findings emphasize the impor- tance of data scaling for improving brain encoding, offering practical insights for optimizing neural network architectures in the context of large-scale stimuli data. This research advances the field of computationally efficient brain encoding, which is crucial for enhancing both computational speed and accuracy. These advance- ments are essential not only for improving our understanding of brain function but also for enabling scalable machine learning models on high-dimensional data and sophisticated stimuli, including applications in neuroprosthetics and clinical neuroscience.