fped-moe-brain-decoding

name	fped-moe-brain-decoding
description	Functional-Network Prior-Guided Mixture-of-Experts (MoE) framework for interpretable brain decoding from fMRI. Uses brain network topology as expert priors with adaptive routing for visual semantic reconstruction.
category	ai_collection
keywords	fMRI decoding, Mixture-of-Experts, brain network, visual reconstruction, functional connectivity, interpretable AI, brain-computer interface
created	2026-05-21T00:00:00.000Z
arxiv_id	2605.19279
arxiv_url	https://arxiv.org/abs/2605.19279
source	cron-neuroscience-research

FPED: Functional-Network Prior-Guided Mixture-of-Experts for Interpretable Brain Decoding

Paper

• Title: FPED: A Functional-Network Prior-Guided Mixture-of-Experts Framework for Interpretable Brain Decoding
• Authors: Yudan Ren, Pengcheng Shi, Zihan Ma, Xiaowei He, Xiao Li
• arXiv: 2605.19279 (2026-05-19)
• URL: https://arxiv.org/abs/2605.19279

Problem

Visual image reconstruction from fMRI is a fundamental brain decoding task for understanding perception and developing BCIs. Current methods flatten fMRI signals from localized visual cortices into 1D vectors and map directly to latent spaces (e.g., CLIP). This paradigm:

1. Disrupts inherent brain network topology
2. Limits neuroscientific interpretability
3. Overlooks synergistic contributions of distributed functional networks in visual semantic processing

Core Methodology: FPED Framework

Architecture

• Functional-Network Priors: Explicitly model different functional brain networks as specialized experts
• Mixture-of-Experts (MoE): Each expert corresponds to a distinct functional brain network (visual, default mode, attention, etc.)
• Adaptive Routing: Learn to route fMRI signals to experts based on semantic content, capturing complementary contributions
• Parameter Efficiency: Only 0.68B parameters yet achieves highly competitive semantic reconstruction

Key Innovations

1. Neurobiologically Grounded Expert Design: Rather than homogeneous MLP layers, each expert maps to a known functional network
2. Structured Representation Learning: Network-level representation preserves brain topology instead of flattening
3. Interpretable Routing Dynamics: Learned routing weights reveal which functional networks contribute to which semantic dimensions
4. Biological Correspondence: Routing patterns show meaningful correspondence between functional networks and modality-specific semantic processing

Technical Approach

fMRI Input → Network-wise Partitioning → Expert Modules (per network)
    ↓
Adaptive Router → Weighted Expert Outputs → CLIP Latent Space → Image Reconstruction

Findings

• FPED achieves competitive semantic reconstruction with only 0.68B parameters
• Routing dynamics reveal biologically meaningful expert specialization: