Overcoming Output Dimension Collapse

This repository contains the code for the paper: “Overcoming Output Dimension Collapse: How Sparsity Enables Zero-shot Brain-to-Image Reconstruction at Small Data Scales” by Kenya Otsuka, Yoshihiro Nagano, and Yukiyasu Kamitani (arXiv).

Overview

The repository consists of two main analysis pipelines:

1. ODC on Real Data (analysis/1_ODC_on_real_data/) - Analysis of ODC using real brain imaging data
2. Sparse Regression (analysis/2_sparse_regression/) - Theoretical and simulation-based analysis of sparse regression

Installation

This project uses uv for dependency management. Install dependencies with:

uv sync

Requirements

• GPU with CUDA support (recommended)
• Sufficient RAM (recommended: 16GB+) for large matrix operations

Quick Start

Dataset Download

Download the required datasets using the provided script:

uv run python scripts/download.py

This will automatically download and organize the following data:

• True features of each dataset in data/features
• fMRI data (.h5) of each dataset in data/fmri
• Model parameters in data/models_shared

The script downloads data from:

• Deep Image Reconstruction@figshare
• brain-decoding-cookbook@figshare
• Spurious reconstruction from brain activity@figshare

ODC on Real Data Analysis

# Step 1: Calculate best prediction
bash analysis/1_ODC_on_real_data/1_calculate_best_prediction/calculate_best_prediction.sh

# Step 2: Calculate brain prediction
bash analysis/1_ODC_on_real_data/2_calculate_brain_prediction/calculate_brain_prediction.sh

# Step 3: Reconstruction
bash analysis/1_ODC_on_real_data/3_reconstruction/iCNN.sh

# Step 4: Plot results
bash analysis/1_ODC_on_real_data/4_plot/calculate-mse.sh

Note: After running the shell script, you can plot the results:

• analysis/1_ODC_on_real_data/4_plot/plot.py - Plot results

Sparse Regression Analysis

# Step 1: Run simulations
bash analysis/2_sparse_regression/1_simulation/simulation.sh

Note: After running the shell script, you can plot the results:

• analysis/2_sparse_regression/2_plot/plot_simulation.py - Plot simulation results
• analysis/2_sparse_regression/2_plot/plot_theory.py - Plot theoretical results

Usage Notes

• All scripts should be run from the project root directory
• The analysis pipelines take significant time to complete
• Use VS Code's Interactive Window for plotting scripts (execute # %% cells)

Project Structure

├── analysis/                                   # Analysis pipelines
│   ├── 1_ODC_on_real_data/                         # Real data ODC analysis
│   └── 2_sparse_regression/                        # Sparse regression analysis
├── src/overcoming_output_dimension_collapse/   # Library
├── assets/                                     # Generated results and data
├── data/                                       # Downloaded data
├── scripts/                                    # Data download script
└── README.md                                   # This file