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. 1FC (first-order functionally-connected) ensembles framework for analyzing information encoding in SNNs through rare coordinated firing events. version: 1.0.0 author: Aditi Aravind, Konstantinos Ladakis, Mario Alexios Savaglio, Stelios M. Smirnakis, Maria Papadopouli arxiv_id: 2606.00073 category: spiking-neural-networks activation_keywords: - functional ensemble - 1FC ensemble - spiking neural network - rare event coding - cofiring analysis - adversarial robustness - functional connectivity - deep SNN created: 2026-06-05

Functional Ensembles as Units of Computation in Deep Spiking Networks

Overview

首次提出 1FC (first-order functionally-connected) ensemble 概念，用于分析深度脉冲神经网络中的信息编码机制。核心发现：信息编码集中在稀有的高协同发放事件中，而非持续性活动。

突破性发现：

SNN 中的功能连接模式与生物皮层相似
1FC ensemble 的集体发放能可靠预测下游响应
信息编码仅在稀有高协同事件时显现
提供精细诊断工具分析信息流

Key Contributions

1. 1FC Ensemble Definition

First-order Functionally-Connected Group:

基于统计显著的 pairwise correlation 形成
来自训练 SNN 的上一层神经元
提供功能连接结构的量化定义

For neuron i in layer L:
1FC(i) = {j in layer L-1 : corr(spike_i, spike_j) > threshold}

2. ReLU-like Input-Output Relationship

Aggregate cofiring predicts downstream response:

1FC ensemble 的集体发放产生可靠的响应预测
输入-输出关系类似 ReLU (thresholded)
Gain 与 ensemble size 系统性相关

Response = ReLU-like(Σ spikes_in_1FC - threshold)
Gain ∝ ensemble_size

3. Rare Event Information Encoding

关键发现：

Reliable class encoding 仅在 高 1FC cofiring 事件 时出现
这些事件本身发生频率很低 (rare but highly coordinated)
信息表示集中在稀有协同模式中

High information content ⊂ Rare high-coordination events

4. Adversarial Robustness Diagnosis

Uniform random noise: 破坏早期和中间层响应
Adversarial perturbations: 扰乱功能连接结构
Weight permutation: 功能连接结构崩溃
提供精细粒度的节点和通路诊断

Technical Framework

Functional Connectivity Analysis

# 计算 pairwise correlation
def compute_1FC_groups(spike_trains_L, spike_trains_L_minus_1, threshold=0.05):
    """
    spike_trains: binary arrays (n_neurons, T)
    """
    correlations = np.corrcoef(spike_trains_L, spike_trains_L_minus_1)
    
    # Statistical significance test
    significant_connections = correlations > threshold
    
    # Form 1FC groups
    1FC_groups = {}
    for i in range(n_neurons_L):
        1FC_groups[i] = np.where(significant_connections[i])[0]
    
    return 1FC_groups

Ensemble Cofiring Analysis

def analyze_cofiring_events(1FC_groups, spike_trains):
    """检测高协同发放事件"""
    
    # Aggregate cofiring for each neuron
    cofiring_strength = []
    for i, ensemble in enumerate(1FC_groups.items()):
        ensemble_spikes = spike_trains[ensemble].sum(axis=0)
        neuron_spike = spike_trains[i]
        
        # Cofiring during neuron's spike
        cofiring_when_spike = ensemble_spikes[neuron_spike > 0]
        cofiring_strength.append(cofiring_when_spike.mean())
    
    # Identify rare high-coordination events
    threshold = np.percentile(cofiring_strength, 95)
    high_cofiring_events = cofiring_strength > threshold
    
    return high_cofiring_events

Information Encoding Detection

def measure_information_encoding(spike_trains, labels, high_cofiring_events):
    """量化稀有事件中的信息编码"""
    
    # Compare class encoding during high vs low cofiring
    spikes_high_cofiring = spike_trains[high_cofiring_events]
    spikes_low_cofiring = spike_trains[~high_cofiring_events]
    
    # Mutual information with class labels
    MI_high = mutual_info_class(spikes_high_cofiring, labels)
    MI_low = mutual_info_class(spikes_low_cofiring, labels)
    
    print(f"Information during high cofiring: {MI_high}")
    print(f"Information during low cofiring: {MI_low}")
    print(f"Ratio: {MI_high / MI_low}")  # >> 1 typically

Implementation Guidelines

When to Use 1FC Analysis

适用场景：

SNN 解释性分析: 理解内部表示如何形成
信息流诊断: 精细粒度分析特定节点/通路
对抗鲁棒性检测: 识别脆弱层和通路
生物神经网络类比: 与皮层功能连接对比

Step-by-Step Workflow

训练 SNN

# Spiking ResNet architecture
model = SpikingResNet(num_layers=5)
model.train_on_dataset(images, labels)

提取 spike trains

# Record all layer spike trains during inference
spike_recordings = {}
for layer_idx in range(num_layers):
    spike_recordings[layer_idx] = model.get_layer_spikes(layer_idx)

形成 1FC groups

# Compute 1FC for each layer
1FC_groups = {}
for L in range(1, num_layers):
    1FC_groups[L] = compute_1FC_groups(
        spike_recordings[L],
        spike_recordings[L-1]
    )

分析 cofiring patterns

# Detect high-coordination events
high_cofiring = analyze_cofiring_events(1FC_groups, spike_recordings)

# Measure information encoding
info_encoding = measure_information_encoding(
    spike_recordings,
    test_labels,
    high_cofiring
)

对抗扰动诊断

# Test adversarial robustness
adversarial_inputs = generate_adversarial(model, test_images)
adversarial_spikes = model.get_all_spikes(adversarial_inputs)

# Compare 1FC structure
1FC_adversarial = compute_1FC_groups(adversarial_spikes)

# Detect disrupted layers
disrupted_layers = compare_1FC_structure(1FC_groups, 1FC_adversarial)

Key Findings Summary

Principle 1: Cortex-like Functional Connectivity

深度 SNN 的功能连接模式与生物皮层观察到的原则一致

Principle 2: Reliable Prediction via Ensemble Cofiring

1FC ensemble 集体发放 → ReLU-like 下游响应，gain 与 size 相关

Principle 3: Concentrated Information in Rare Events

稀有高协同事件集中了大部分信息编码能力

Principle 4: Learning Shapes Connectivity

学习过程塑造功能连接结构，weight permutation 破坏这种结构

Principle 5: Targeted Diagnostics

允许在特定节点和通路进行精细粒度诊断

Experimental Results

Rare Event Statistics

High 1FC cofiring events frequency: < 5% of total spikes
Information encoding during high cofiring: >> during low cofiring
Ensemble size range: typically 5-20 neurons

Layer-wise Disruption Patterns

Early layers: More sensitive to uniform noise
Intermediate layers: Most affected by adversarial perturbations
Late layers: Relatively robust but critical for output

Biological Correlation

1FC patterns resemble cortical columnar organization
Similar to observed ensembles in motor cortex
Consistent with rare but coordinated firing in biology

Pitfalls & Common Mistakes

1. Threshold Selection

❌ 使用固定阈值不进行统计检验
✅ 使用 statistical significance test (e.g., permutation test)

2. Ensemble Size Ignorance

❌ 只看 cofiring 不考虑 ensemble size
✅ 同时分析 ensemble size 和 cofiring strength

3. Layer Selection Bias

❌ 只分析最后一层
✅ 分析所有层，早期层往往最脆弱

4. Temporal Window Oversimplification

❌ 使用过大时间窗口合并 spikes
✅ 使用合适的 temporal precision (e.g., 1-5ms bins)

5. Ignoring Baseline

❌ 不比较 random vs trained network
✅ 对比 baseline 以确认学习塑造的连接

Comparison with Related Work

Approach	Scale	Temporal	Interpretability	Biological Plausibility
1FC Ensemble	Fine	High	Excellent	High
Simple Correlation	Fine	Low	Medium	Medium
PCA Clustering	Coarse	None	Low	Low
Graph Neural	Coarse	None	Medium	Medium

Extensions & Applications

1. Multi-layer Extension

扩展到 2FC, 3FC (higher-order functional connectivity)

2. Temporal Dynamics

分析 1FC ensemble 的时序演化

3. Cross-architecture Comparison

比较不同 SNN architecture 的 1FC patterns

4. Biological Data Validation

与实际皮层神经记录数据对比

5. Targeted Intervention

基于诊断结果设计 targeted fine-tuning

Code Resources

Spiking ResNet implementation: [spiking-resnet]
Correlation analysis: [numpy.corrcoef], [scipy.stats]
Adversarial attack generation: [foolbox], [advertorch]

References

Aravind A, Ladakis K, Savaglio MA, Smirnakis SM, Papadopouli M. "Rare Events, Real Signals: Functional Ensembles as Units of Computation in Deep Spiking Networks" arXiv:2606.00073
Related work on cortical ensembles: [citations needed]
Spiking neural network foundations: [gerstner2014]

Created: 2026-06-05 Source: arXiv:2606.00073