| name | dual-axis-zebrafish-circuits |
| description | 斑马鱼被盖微电路双轴归因方法论。基于斑马鱼视网膜被盖微电路的双轴功能归因(能量高效信息处理和鲁棒稳定)方法,将生物学电路组织转化为类脑神经网络架构设计。适用于脉冲神经网络、类脑架构设计、电路归因。触发词:dual-axis, 双轴归因, zebrafish tectal, microcircuit attribution, energy efficiency, robustness, ns_TIN, superficial_TIN |
Dual-Axis Zebrafish Tectal Microcircuit Attribution Methodology
斑马鱼被盖微电路的双轴功能归因方法,将生物学电路结构转化为类脑神经网络设计模式。
Source Paper
Title: Dual-axis attribution of zebrafish tectal microcircuits for energy-efficient and robust neurocomputing
arXiv: 2605.13924
Date: 2026-05-15
Authors: Ningping Li, Hao Zhang, Yi Zhou
Category: cs.NE (Neural and Evolutionary Computing)
URL: https://arxiv.org/abs/2605.13924
Core Concept
生物神经网络包含支持不同计算功能的特化子结构,但许多类脑神经网络借用生物基序而不识别其电路级起源。本方法提出沿两个计算轴对斑马鱼被盖微电路进行功能归因:
- Energy-Efficient Information Processing(能量高效信息处理)
- Robustness-Preserving Stabilization(鲁棒保存稳定)
Key Innovations
1. Directed Retinotectal Microcircuit Graph Reconstruction
- 从斑马鱼视网膜被盖系统重建有向微电路图
- 通过动态仿真验证视网膜被盖信号传播
- 建立从生物结构到计算图的映射
2. Dual-Axis Attribution Framework
通过 LIF (Leaky Integrate-and-Fire) 脉冲神经网络作为非线性扰动测试台:
Energy Sensitivity Index (ESI)
$$ESI_i = \frac{|E_{baseline} - E_{ablated}|}{E_{baseline}}$$
- 测量子电路 $i$ 被消融后的能量敏感性
- 低 ESI 但显著影响预测误差 → 能量高效信息门
Robustness Sensitivity Index (RSI)
$$RSI_i = \frac{|R_{baseline} - R_{ablated}|}{R_{baseline}}$$
- 测量子电路 $i$ 被消融后的鲁棒性敏感性
- 高 RSI → 反馈式系统稳定性维护
3. Functional Dissociation Findings
| Subcircuit | Spike Footprint | Sensitivity | Role |
|---|
| ns_TIN | Low | Prediction error (ESI) | Spike-efficient internal information gate |
| superficial_TIN | Higher | System robustness (RSI) | Feedback-like stability maintenance |
4. Transfer to Artificial Neural Networks
将归因功能转移到 ResNet18 架构中:
- ns_TIN-inspired module: 在推理预算缩减下提升性能保持
- superficial_TIN-inspired module: 在输入噪声下提升鲁棒性
Methodology Steps
Step 1: Microcircuit Reconstruction
import networkx as nx
def build_retinotectal_graph():
"""Construct directed microcircuit graph from zebrafish tectal connectivity."""
G = nx.DiGraph()
G.add_nodes_from([
'RGC',
'ns_TIN',
'superficial_TIN',
'TPC',
'Pyr',
])
G.add_edges_from([
('RGC', 'ns_TIN'),
('RGC', 'superficial_TIN'),
('ns_TIN', 'TPC'),
('superficial_TIN', 'TPC'),
('superficial_TIN', 'ns_TIN'),
])
return G
Step 2: LIF-SNN Ablation Testing
def ablation_test(G, target_subcircuit, model):
"""Selectively ablate a subcircuit and measure sensitivity."""
baseline_pred = model.predict(G)
baseline_energy = compute_spike_footprint(model)
baseline_robust = evaluate_robustness(model, noise_level=0.1)
G_ablated = G.copy()
G_ablated.remove_nodes_from(target_subcircuit)
ablated_pred = model.predict(G_ablated)
ablated_energy = compute_spike_footprint(model)
ablated_robust = evaluate_robustness(model, noise_level=0.1)
ESI = abs(baseline_energy - ablated_energy) / baseline_energy
RSI = abs(baseline_robust - ablated_robust) / baseline_robust
return {'ESI': ESI, 'RSI': RSI}
Step 3: Functional Transfer
class DualAxisModule(nn.Module):
"""Dual-axis inspired neural architecture module."""
def __init__(self, in_channels, out_channels, mode='energy'):
super().__init__()
self.mode = mode
if mode == 'energy':
self.gate = nn.Sequential(
nn.Linear(in_channels, in_channels // 4),
nn.ReLU(),
nn.Linear(in_channels // 4, in_channels),
nn.Sigmoid()
)
elif mode == 'robustness':
self.stabilizer = nn.Sequential(
nn.LayerNorm(in_channels),
nn.Linear(in_channels, in_channels),
nn.GELU(),
nn.Dropout(0.1),
)
def forward(self, x):
if self.mode == 'energy':
gate = self.gate(x.mean(dim=-1))
return x * gate.unsqueeze(-1)
else:
return self.stabilizer(x) + x
Design Patterns
Pattern 1: Energy-Efficient Information Gate (ns_TIN)
Input → [Lightweight Gate] → Sparse Activation → Output
↓
Computation Budget
Reduction Module
Use when: Need to reduce computation while preserving key information flow
Pattern 2: Robustness Stabilizer (superficial_TIN)
Input → [Normalization] → [Feedback Path] → [+ Residual] → Output
↓ ↓
Stabilizer Robustness
Module Enhancement
Use when: Need to improve robustness against noise/perturbations
Application Guidelines
-
Architecture Design: Use dual-axis modules in hybrid networks
- Front layers: ns_TIN for efficient early processing
- Deep layers: superficial_TIN for robust feature stabilization
-
Resource-Constrained Deployment: ns_TIN-inspired gating for edge deployment
-
Safety-Critical Systems: superficial_TIN-inspired stabilization for robust operation
Verification
- Test ESI/RSI tradeoff on target architecture
- Validate on CIFAR-10 with budget reduction and noise corruption
- Compare spike footprint vs. accuracy preservation
- Measure robustness improvement under adversarial conditions
Related Skills
spiking-neural-network-analysis
brain-inspired-snn-pattern-analysis
brain-inspired-intelligence-paradigm
energy-regularized-neural-mpc
Activation Keywords
- dual-axis attribution
- 双轴归因
- zebrafish tectal microcircuit
- microcircuit attribution
- energy efficient neurocomputing
- robustness-preserving stabilization
- ns_TIN
- superficial_TIN
- retinotectal circuit
- biological circuit transfer
- LIF-SNN ablation
- circuit-level bio-inspiration