This project reproduces a subset of the analysis conducted in A High-Resolution 7-Tesla fMRI Dataset from Complex Natural Stimulation with an Audio Movieby Hanke et al. We analyze inter-subject correlation among 5 subjects, then apply machine learning to see if we can predict if a subject was listening to an interior or exterior scene of the movie based on brain state. The results for our reproduction are consistent with those produced by Hanke et al. and our internal-external scene classifier has a validation accuracy of 0.942.

The raw data is approximately 37.5 GBs. Additionally, the complete analysis will - in total - produce another approximately 190 GBs of data written to disk. Please ensure your machine is equippted with sufficient storage space.

Since large matrices must be computed and held in memory, systems should have at least 90 GBs of accessible RAM. We strongly encourage running on a machine with 120 GBs of accessible RAM to emulate development environment.

1. Clone and move into the project repository: git clone https://github.com/berkeley-stat159/project-lambda.git && cd project-lambda
2. Install python dependencies with pip: pip install -r requirements.txt
3. Code depends on an internally provided python module module called stat159lambda. Ensure this module is on your python path with export PYTHONPATH='<absolute-path-to-repository>/project-lambda/code'.
   For example, export PYTHONPATH='/Users/alondaks/project-lambda/code'. We strongly encourage adding your export statement to your .bashrc while running this project to ensure the environement variable is defined in all shells.

To run python unit tests: make test from the top level project directory.

Please run the follow make commands in order from top to bottom from within the top level project directory.

1. Download and verify data

• make data fetches and downloads raw data
• make validate-data verifies data integrity against known checksums

2. Preprocess the data

• make preprocess concatenates runs, applies guassian filter, and reshapes data to 2d

3. Exploratory data analysis

• make eda produces eda plots and simulations

4. Reproduction

• make reproduction runs all correlation reproduction work and produces analysis figures

5. Classification

• make classification-cross-validate runs random forrest cross validation and saves cv accuracies to file
• make classification-validation tests a fully trained random forrest on validation set and saves accuracy to file

To run all analysis excluding data download and verification (steps 2-5 only): make all-analysis

To generate the PDF report run make generate-paper from the top level project directory. To remove remove intermediary latex files run make clean-paper

• Alon Daks (alondaks)
• Jordeen Chang (jodreen)
• Lisa Ann Yu (lisaannyu)
• Ying Luo (yingtluo)

Special thank you to Jarrod Millman, Matthew Brett, Ross Barnowski, and J-B Poline for their continued support, instruction and encouragment with this project