| name | dual-memory-pathway-snn |
| description | Dual Memory Pathway (DMP) neuromorphic network co-design methodology. Inspired by cortical fast-slow organization, combines explicit slow memory with fast spiking activity for long-timescale context. Applies to: neuromorphic computing, event-driven sensing, energy-efficient SNN deployment. Activation: dual memory pathway, neuromorphic co-design, fast-slow SNN, near-memory compute, cortical memory. |
Dual Memory Pathway (DMP) Neuromorphic Networks
Algorithm-hardware co-design combining biological fast-slow cortical organization with neuromorphic hardware optimization for energy-efficient long-timescale computation.
Metadata
- Source: arXiv:2512.07602
- Authors: Pengfei Sun, Zhe Su, Jascha Achterberg, Giacomo Indiveri, Dan F.M. Goodman, Danyal Akarca
- Published: 2025-12-08 (v2: 2025-12-11)
Core Methodology
Key Innovation
The DMP architecture introduces an explicit slow memory pathway alongside fast spiking activity, inspired by cortical fast-slow organization. Each layer maintains a compact low-dimensional state summarizing recent activity that modulates spiking dynamics.
Algorithm Level
- Dual Memory Structure: Each layer has two pathways:
- Fast pathway: Standard spiking activity (event-driven, sparse)
- Slow pathway: Compact state vector tracking long-timescale context
- Memory Modulation: Slow pathway state modulates fast spiking dynamics
- Learning Stabilization: Explicit memory stabilizes training while preserving sparsity
- Parameter Efficiency: 40-60% fewer parameters than equivalent SOTA SNNs
Hardware Level
- Near-Memory-Compute Architecture: Retains compact shared state while optimizing dataflow
- Heterogeneous Dataflow: Optimized for both sparse-spike and dense-memory pathways
- Throughput Improvement: 4x increase over state-of-the-art implementations
- Energy Efficiency: 5x improvement in energy efficiency
Technical Framework
DMP Layer Structure
For each layer l:
Fast state: s_l(t) = spike_activation(input, weights, slow_state)
Slow state: m_l(t+1) = α·m_l(t) + (1-α)·aggregate(s_l(t))
Output: modulated_spike(s_l(t), m_l(t))
Where:
- α is the memory decay factor (slow timescale)
- aggregate() summarizes fast activity into slow state
- modulation() combines fast and slow pathways
Implementation Steps
- Define DMP layer with dual pathway structure
- Implement slow state update with configurable timescale
- Design modulation mechanism (multiplicative, additive, or gating)
- Co-design hardware dataflow for heterogeneous pathways
- Train with standard SNN surrogate gradient + memory regularization
Applications
- Event-based vision sensors with long-timescale context
- Neuromorphic edge computing for autonomous systems
- Low-power IoT sensing with temporal memory
- Real-time neuromorphic computation and learning
Pitfalls
- Memory decay factor α must be tuned per task
- Hardware co-design requires specialized architecture knowledge
- Slow pathway introduces additional compute overhead
- Not suitable for purely feedforward, single-pass inference
Related Skills
- dual-timescale-memory-spiking-neuron-astrocyte
- snn-universal-approximation-theory
- snn-edge-intelligence-survey
- physical-foundation-models