Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ChronoPlastic Spiking Neural Networks

Sarim Chaudhry

TL;DR

This work addresses the difficulty of learning long-range temporal dependencies in spiking neural networks due to fixed time constants. It proposes ChronoPlastic Spiking Neural Networks (CPSNNs), where each synapse maintains fast and slow traces and modulates the slow decay with a learned warp factor, enabling input-conditioned memory without external memory or attention. Trained end-to-end with surrogate gradients, CPSNNs achieve faster convergence and higher accuracy on long-horizon tasks while preserving linear-time, neuromorphic-compatible computation. The results suggest adaptive temporal modulation at the synaptic level as a key mechanism for scalable temporal learning in spiking systems, with potential for efficient hardware deployment and broader temporal cognition applications.

Abstract

Spiking neural networks (SNNs) offer a biologically grounded and energy-efficient alternative to conventional neural architectures; however, they struggle with long-range temporal dependencies due to fixed synaptic and membrane time constants. This paper introduces ChronoPlastic Spiking Neural Networks (CPSNNs), a novel architectural principle that enables adaptive temporal credit assignment by dynamically modulating synaptic decay rates conditioned on the state of the network. CPSNNs maintain multiple internal temporal traces and learn a continuous time-warping function that selectively preserves task-relevant information while rapidly forgetting noise. Unlike prior approaches based on adaptive membrane constants, attention mechanisms, or external memory, CPSNNs embed temporal control directly within local synaptic dynamics, preserving linear-time complexity and neuromorphic compatibility. We provide a formal description of the model, analyze its computational properties, and demonstrate empirically that CPSNNs learn long-gap temporal dependencies significantly faster and more reliably than standard SNN baselines. Our results suggest that adaptive temporal modulation is a key missing ingredient for scalable temporal learning in spiking systems.

ChronoPlastic Spiking Neural Networks

TL;DR

This work addresses the difficulty of learning long-range temporal dependencies in spiking neural networks due to fixed time constants. It proposes ChronoPlastic Spiking Neural Networks (CPSNNs), where each synapse maintains fast and slow traces and modulates the slow decay with a learned warp factor, enabling input-conditioned memory without external memory or attention. Trained end-to-end with surrogate gradients, CPSNNs achieve faster convergence and higher accuracy on long-horizon tasks while preserving linear-time, neuromorphic-compatible computation. The results suggest adaptive temporal modulation at the synaptic level as a key mechanism for scalable temporal learning in spiking systems, with potential for efficient hardware deployment and broader temporal cognition applications.

Abstract

Spiking neural networks (SNNs) offer a biologically grounded and energy-efficient alternative to conventional neural architectures; however, they struggle with long-range temporal dependencies due to fixed synaptic and membrane time constants. This paper introduces ChronoPlastic Spiking Neural Networks (CPSNNs), a novel architectural principle that enables adaptive temporal credit assignment by dynamically modulating synaptic decay rates conditioned on the state of the network. CPSNNs maintain multiple internal temporal traces and learn a continuous time-warping function that selectively preserves task-relevant information while rapidly forgetting noise. Unlike prior approaches based on adaptive membrane constants, attention mechanisms, or external memory, CPSNNs embed temporal control directly within local synaptic dynamics, preserving linear-time complexity and neuromorphic compatibility. We provide a formal description of the model, analyze its computational properties, and demonstrate empirically that CPSNNs learn long-gap temporal dependencies significantly faster and more reliably than standard SNN baselines. Our results suggest that adaptive temporal modulation is a key missing ingredient for scalable temporal learning in spiking systems.
Paper Structure (65 sections, 5 theorems, 48 equations, 5 figures, 1 table, 1 algorithm)

This paper contains 65 sections, 5 theorems, 48 equations, 5 figures, 1 table, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Temporal Learning in Spiking Neural Networks
  4. Adaptive Time Constants
  5. Attention and External Memory
  6. Our Position
  7. ChronoPlastic Spiking Neural Networks
  8. Neuron Model
  9. ChronoPlastic Synapse
  10. Synaptic Current
  11. Training and Optimization
  12. Experiments
  13. Long-Gap Temporal XOR Task
  14. Baselines
  15. Results
  16. ...and 50 more sections

Key Result

Theorem 1.1

Consider the CPSNN slow trace $z_t = \alpha_s^{\omega_t} z_{t-1} + s_t$ with $\omega_t \in (0,1)$, and the fixed-decay trace $\tilde{z}_t = \alpha_s \tilde{z}_{t-1} + s_t$. Then: (i) Stationarity gap. If $\omega_t$ is not almost surely constant in time (i.e., there exist $t\neq t'$ with $\omega_t \n with strict inequality $\kappa_{t,k} > \alpha_s^{t-k}$ whenever $\omega_j < 1$ for at least one $j

Figures (5)

  • Figure 1: Training accuracy over epochs on the long-gap temporal XOR task. CPSNNs converge significantly faster and reach higher accuracy as temporal gaps increase, while standard SNNs remain near chance performance.
  • Figure 2: Learned warp factor $\omega_t$ over time. The model selectively reduces $\omega_t$ around informative cue events, slowing decay and preserving memory, while increasing $\omega_t$ during distractors.
  • Figure 3: Gradient magnitude across time steps during training. Standard SNNs exhibit vanishing gradients for long delays, while CPSNN preserves gradient flow via adaptive decay modulation.
  • Figure 4: Local synaptic computation in CPSNNs versus global temporal interactions in attention-based models. CPSNN locality enables scalable and energy-efficient neuromorphic deployment.
  • Figure 5: Memory usage as a function of sequence length. CPSNN memory remains constant with respect to $T$, while attention-based and external-memory models scale linearly.

Theorems & Definitions (10)

  • Theorem 1.1: Strict non-stationarity and controllable memory horizon
  • proof
  • Proposition 1.2: CPSNN traces realize non-stationary exponential kernels unattainable by any fixed-decay trace
  • proof
  • Corollary 1.3
  • proof
  • Theorem 1.4
  • proof
  • Theorem 3.1
  • proof