spiking-neural-networks-fmri-visual-decoding

name	spiking-neural-networks-fmri-visual-decoding
title	Spiking Neural Networks for fMRI-Based Visual Semantic Decoding
version	1.0.0
description	Methodology for using Spiking Neural Network (SNN)-derived visual features as targets for fMRI-based visual semantic decoding, showing superior alignment with brain activity compared to traditional ANN features.
trigger_words	["snn fmri decoding","spiking neural network brain decoding","fMRI visual semantic decoding","brain-decodable visual representations"]
domain	neuroscience/computational-neuroscience
authors	["Jiahong Zhang","Jinning Zhao","Sijun Shen","Siyuan Xu","Bo Xu","Guoqi Li"]
paper_id	arXiv:2607.19170
date	2026-07-21T00:00:00.000Z

Spiking Neural Networks for fMRI-Based Visual Semantic Decoding

Overview

This methodology investigates Spiking Neural Network (SNN)-derived visual features as alternative targets for fMRI-based visual semantic decoding. The research demonstrates that SNN-derived features exhibit stronger alignment with fMRI responses and significantly improve visual semantic decoding performance compared to traditional Artificial Neural Network (ANN) baseline features.

Key Findings

Performance Improvements

• Feature-prediction error reduced from 0.7707 (ANN) to 0.0282 (SNN)
• Top-1 semantic decoding accuracy improved from 0.1800 (ANN) to 0.4400 (SNN) on the GoD dataset
• Both spiking neural dynamics and temporal simulation steps contribute to the observed advantage

Methodological Approach

• Uses the same L2-regularized linear fMRI-to-feature decoder across all models
• Only varies the feature vectors used as regression targets
• Compares ANN baseline with four SNN variants from the same architectural family
• SNN variants differ in their spiking dynamics while maintaining architectural consistency

Implementation Steps

1. Model Selection and Training

• Select SNN architecture from the same family as your ANN baseline
• Ensure SNN variants have different spiking dynamics (e.g., different neuron models, time constants)
• Train SNN models on the same visual dataset as the ANN baseline

2. Feature Extraction

• Extract visual features from SNN models at the appropriate layer(s)
• For temporal SNNs, consider features across multiple time steps or aggregate temporal information
• Normalize features consistently with ANN baseline for fair comparison

3. fMRI-to-Feature Mapping

• Use L2-regularized linear regression to map fMRI responses to feature vectors
• Apply the same regularization parameters across ANN and SNN targets
• Validate mapping performance using cross-validation

4. Semantic Decoding Evaluation

• Evaluate downstream semantic decoding performance using standard metrics
• Compare against ANN baseline using identical evaluation protocols
• Perform ablation studies to isolate contributions of spiking dynamics vs. temporal simulation

Best Practices

Target Feature Design