Testing the Brain Alignment of Models with High Effective Dimensionality

Basic Overview

This project investigates the relationship between effective dimensionality(ED) and brain alignment in the vision models using the framework Brain-Score. Previous studies have highlighted both the benefits of low and high dimensionality in deep neural networks (DNNs).

Specifically, a recent study found a potential connection between high ED and improved alignment with brain activity. This project aims to test the findings in the recent study by expading the scope of layers and models for investigation. Additionally, we aim to offer new perspectives on model interpretability, brain-inspired AI, and the broader connection between artificial and biological intelligence.

This project is conducted as a project2 of the CS-433 ML course Fall 2024 at EPFL. collaborating with neuroAI lab in EPFL.

Note: The report slightly exceeds 4 pages as it was not possible to adjust the spacing within the document. With appropriate space adjustments, all the content would fit within 4 pages.

File Configuration

• data
  • Directory containing the datasets used for analysis.
• jobs
  • Directory containing shell script used for submitting jobs to cluster.
• logs
  • Directory containing logs obtained from izar job submission.
• notebooks
  • Contains jupyter notebook for extracting features from models in interest.
• resources
  • Contains resources for analysis
• ed-calculation.ipynb
  • Jupyter notebook for calculating effective dimensionality(ED) based on extracted features from the layers vision models.
• extraction-cat.ipynb
  • Extracts feature from image and draws feature map
• feature-map.ipynb
  • Creates feature map from the feature extraction
• layer1.png
  • Layer1 feature map created with extraction-cat.ipynb
• plots.ipynb
  • Contains code for reproducing plots in the report
• reshape.ipynb
  • Jupyter notebook for reshaping the xarray formed feature extraction to numpy array.
• wrong-ed-calculation.ipynb
  • Contains wrong ed calculation code, calculating singular separately. Final ed-calculation.ipynb contains the correct version.

Execution Requirements

1. Clone the repository

2. Install the required dependencies: pip install -r requirements.txt

Team information

name	sciper
Hamza Remmal	310917
Lina Sadgal	342075
Ahyoung Seo	390238