How kamitanilab/genericobjectdecoding works
Project Type and Purpose
• This is an academic/demo repository for “Generic Object Decoding” from fMRI data using hierarchical visual features.
• It provides data organization and two parallel analysis pipelines (Matlab and Python) that replicate experiments from Horikawa & Kamitani (2017).
File‐Structure Analysis
• Top‐level:
• code/ (contains two subfolders matlab/ and python/)
• data/ (holds raw & preprocessed fMRI data references, stimulus tables, CNN model definitions)
• docs/ (GitHub Pages site content)
• code/matlab/ uses:
• analysis_*.m scripts for Category Identification, Feature Prediction, Accuracy
• conversion, figure‐export, volume‐export utilities
• lib/ contains two major libraries: SPR_2009_12_17 (sparse regression tools) and mex_prog_2010‐06‐21 (compiled MEX functions)
• code/python/ implements equivalent analyses in Python, plus:
• god_config.py for global settings
• makedata_features.py to assemble feature datasets
• conversion scripts to import Matlab results
• createfigure.py to reproduce publication figures
• data/ holds:
• stimulus tables (.tsv/.csv) and category mappings
• CNN model definitions (Caffe prototxt under data/cnn)
• ImageNet test/training lists and related metadata
• README files in each directory describe dependencies and usage.
Architecture Insights from README
• External data sources: OpenNeuro (raw fMRI), Figshare (preprocessed fMRI & image features), private stimulus‐image requests.
• Both pipelines expect the same base data schema and converge on figure‐generation.
Main Components
a. Data Layer
• External repositories: OpenNeuro, Figshare, Stimulus‐image server (request form)
• Local data folder: stimulus tables, image URLs, CNN prototxt, category label files
b. Processing Layer
• Matlab Pipeline
– SPR sparse‐regression submodule
– MEX accelerated functions (mex_prog)
– Analysis scripts → decoded feature outputs
• Python Pipeline
– Data‐preparation (makedata_features)
– Analysis scripts mirroring Matlab logic
– Merge & postprocess results, plus figure routines
c. Model Layer
• Hierarchical feature extractor: Caffe “bvlc_reference_caffenet” (data/cnn)
• Feature representations (fc6 activations, etc.)
d. Output Layer
• Decoding results (CSV/MAT files)
• Publication‐quality figures (PNG/PDF)
• Logs & intermediate result files
Relationships and Data Flow
• Stimulus tables + fMRI data → both Matlab & Python pipelines
• Caffe model definitions → feature‐extraction subroutines → analysis scripts
• Matlab outputs → conversion script for Python → merged analysis
• Final outputs from either pipeline → createfigure modules → figure files
Architectural Patterns & Design Principles
• Modular, language‐parallel pipelines share the same data schema
• Separation of concerns:
• Data ingestion & conversion
• Core decoding algorithms (SPR library / custom scripts)
• Visualization & export
• Reuse of established libraries (SPR_2009, MEX modules, Caffe)
• Extensibility: additional analysis scripts can be plugged into either pipeline
Elements
• Clear component names: “SPR_2009 Library,” “MEX Prog Module,” “makedata_features.py,” “Caffe prototxt”