functional-ensembles-deep-spiking-networks - SKILL.md Agent Skill

name: functional-ensembles-deep-spiking-networks description: "Functional Ensembles as Units of Computation in Deep Spiking Networks. 分析深度脉冲神经网络中功能连接组的计算单元作用，通过一阶功能连接（1FC）组揭示信息编码机制。Activation: functional ensemble, 1FC group, spiking neural network, functional connectivity, information encoding, deep SNN, rare events cofiring."

Functional Ensembles as Units of Computation in Deep Spiking Networks

Paper Information

arXiv ID: 2606.00073
Title: Rare Events, Real Signals: Functional Ensembles as Units of Computation in Deep Spiking Networks
Authors: Aditi Aravind, Konstantinos Ladakis, Mario Alexios Savaglio, Stelios M. Smirnakis, Maria Papadopouli
Categories: cs.NE, cs.AI, cs.LG
Submitted: 21 May 2026
DOI: https://doi.org/10.48550/arXiv.2606.00073

Core Contributions

1. First-Order Functionally-Connected (1FC) Groups

引入一阶功能连接组的概念，定义为基于与前一层神经元统计显著的成对相关性：

1FC Group Definition:
- Neuron's statistically significant pairwise correlations
- With neurons from previous layer in trained SNN
- Based on functional connectivity principles from biological cortex

2. Ensemble Cofiring Properties

关键发现：

1FC 组的聚合协同发射可靠预测下游神经元响应
ReLU-like 输入-输出关系，增益随 ensemble size 系统性变化
信息编码仅在高 1FC 协同发射事件中出现
协同发射事件本身发生频率低（稀有事件）

3. Robustness Analysis

在两种扰动条件下测试：

Uniform Random Noise: 早期和中间层的响应模式被破坏
Adversarial Perturbations: 功能连接结构被显著扰乱

4. Learning-Dependent Structure

证明：

功能连接结构由学习过程塑造
权重置换（weight permutation）会破坏该结构
证实 1FC ensembles 是功能上有意义的计算基元

Key Methodology

Analysis Framework

# 1FC Ensemble Formation
def form_1fc_group(neuron, prev_layer_neurons):
    """
    Form first-order functionally-connected group
    based on statistical pairwise correlations
    """
    correlations = compute_pairwise_correlations(neuron, prev_layer_neurons)
    significant_pairs = statistical_test(correlations, threshold=0.05)
    return significant_pairs

# Ensemble Cofiring Analysis
def analyze_cofiring(1fc_ensemble, spike_trains):
    """
    Track response properties during inference
    """
    cofiring_events = detect_cofiring(spike_trains)
    downstream_response = measure_downstream_activity(cofiring_events)
    return cofiring_response_relationship

Information Encoding Pattern

Rare Event Encoding:
- High 1FC cofiring events → Reliable encoding of presented class
- Low frequency of occurrence → Sparse but highly informative
- Concentrated information in coordinated activity patterns

Applications

1. Fine-Grained Diagnostics

用于信息流的高分辨率诊断：

目标节点和通路检查
噪声和对抗扰动的脆弱性分析
层级信息传递质量评估

2. Architecture Design Insights

功能连接驱动的网络设计
Ensemble-based 信息传递优化
稀有事件编码的鲁棒性增强

3. Neuroscience-Inspired Analysis

将系统神经科学概念应用于 SNN：

功能连接图谱分析
Ensemble 协同活动追踪
生物合理性验证

Implementation Guidelines

Step 1: Train SNN Architecture

Recommended: Spiking ResNet
- Hierarchical processing structure
- Train with surrogate gradient method
- Achieve task performance baseline

Step 2: Compute Functional Connectivity

# Pairwise correlation computation
from scipy.stats import pearsonr

def compute_fc_matrix(layer_activations):
    fc_matrix = np.zeros((n_neurons, n_neurons_prev))
    for i in range(n_neurons):
        for j in range(n_neurons_prev):
            fc_matrix[i,j] = pearsonr(
                layer_activations[i],
                prev_layer_activations[j]
            )[0]
    return fc_matrix

Step 3: Identify 1FC Ensembles

def identify_1fc_ensembles(fc_matrix, threshold=0.05):
    """
    Extract statistically significant functional groups
    """
    from scipy.stats import threshold_check
    
    ensembles = {}
    for neuron_id in range(fc_matrix.shape[0]):
        significant_connections = threshold_check(
            fc_matrix[neuron_id],
            alpha=threshold
        )
        ensembles[neuron_id] = significant_connections
    
    return ensembles

Step 4: Cofiring Event Detection

def detect_cofiring_events(spike_trains, ensemble_members):
    """
    Detect rare high-cofiring events
    """
    cofiring_threshold = len(ensemble_members) * 0.7
    
    cofiring_events = []
    for time_step in range(spike_trains.shape[1]):
        active_count = sum(
            spike_trains[m][time_step] for m in ensemble_members
        )
        if active_count >= cofiring_threshold:
            cofiring_events.append(time_step)
    
    return cofiring_events

Key Results Summary

Metric	Finding
Ensemble Cofiring Response	ReLU-like, gain scales with ensemble size
Encoding Reliability	High only during rare cofiring events
Noise Robustness	Disrupted in early/intermediate layers
Weight Permutation	Breaks functional connectivity structure
Learning Dependency	Structure shaped by training process

Potential Pitfalls

1. Threshold Selection

避免使用固定阈值
建议：根据网络深度和任务动态调整
早期层需要更宽松的显著性标准

2. Sparse Activity Handling

稀有事件统计需要足够样本
建议：长时间推理窗口收集数据
或：使用 bootstrap 方法增强统计可靠性

3. Layer-Specific Analysis

不同层功能连接模式差异大
建议：分层独立分析
避免跨层统一阈值

Activation Keywords

functional ensemble
1FC group
spiking neural network
functional connectivity
information encoding
deep SNN
rare events cofiring
ensemble computation
neural ensemble analysis

Recommended Model

sonnet4.5: 适合方法实现和实验设计
opus4.5: 适合深入理论分析和神经科学背景

Related Skills

spiking-neural-network-analysis: SNN 论文分析框架
brain-graph-neural: 脑网络图神经网络方法
neural-population-dynamics: 神经种群动力学分析
functional-connectivity-analysis: 功能连接分析方法

References

Original Paper: arXiv:2606.00073
Functional Connectivity in Cortex: Neuroscience literature on functional ensembles
Information Theory: Shannon entropy and mutual information for neural coding
SNN Training: Surrogate gradient methods and temporal coding

Future Directions

Extension to Higher-Order FC: 2FC, 3FC groups for multi-layer interactions
Temporal Dynamics: Time-varying functional connectivity analysis
Adversarial Defense: Ensemble-based robust encoding strategies
Hardware Implementation: Neuromorphic deployment of ensemble-based architectures