| name | local-pheromone-network |
| category | ai_collection |
| trigger_words | ["local pheromone network","pheromone-weighted learning","synaptic traces","consolidation replay","structural plasticity","partitioned memory","信息素网络","局部学习"] |
| description | Local Pheromone Network methodology for sparse, local, manually updated neural networks without backpropagation. Uses pheromone-weighted Hebbian updates with short-term/long-term synaptic traces, consolidation, and replay. Achieves partitioned memory preservation, conflict reduction, and structural plasticity through biologically-inspired mechanisms.
|
| arxiv_id | 2606.30669 |
| authors | ["Xingcheng Fu","Xianjun Chen","Zhihao Li"] |
| affiliation | Not specified |
| date | 2026-06-22 |
Local Pheromone Network: Sparse Local Learning with Multi-Scale Synaptic Trails
Overview
Local Pheromone Network is a research prototype for sparse, local, manually updated neural networks that operates without automatic differentiation. Each output unit reads only a fixed local neighborhood of input units subject to geometric distance and molecular-tag compatibility.
The system implements biologically-inspired mechanisms:
- Short-term pheromone traces: Immediate synaptic activity markers
- Long-term pheromone traces: Consolidated synaptic strength indicators
- Consolidation state: Optional mechanism for memory stabilization
- Structural plasticity: Dynamic connectivity changes
- Local replay: Reactivation of recent patterns for conflict resolution
Core Methodology
Local Connectivity Constraints
Each output unit has:
- Fixed local neighborhood: Reads only nearby input units
- Geometric distance constraint: Connection probability decays with distance
- Molecular-tag compatibility: Synapses must match molecular markers (inspired by synaptic tagging)
This contrasts with fully-connected networks where every output can read every input.
Pheromone-Weighted Hebbian Updates
Training does not use backpropagation. Instead, each layer performs:
- Pheromone-weighted update: Synaptic changes weighted by pheromone trace strength
- Budgeted subset selection: Only a subset of local synapses updated per timestep
- Local error + co-activity: Selection based on local error signal and pre/post co-activation
The update budget adapts online:
- Shrinks when loss improves: Focus on currently successful synapses
- Expands when loss worsens: Explore new synaptic neighborhoods
Multi-Scale Trace System
Each synapse stores:
- Weight: Current synaptic strength
- Short-term pheromone trace: Recent activity (fast decay)
- Long-term pheromone trace: Consolidated strength (slow decay)
- Consolidation state: Optional flag for stabilized memories
The dual-trace system separates:
- : Short-term trace mediates rapid changes
