Multi-Timescale Conductance Spiking Networks: A Sparse, Gradient-Trainable Framework with Rich Firing Dynamics for Enhanced Temporal Processing
Multi-Timescale Conductance Spiking Networks: A Sparse, Gradient-Trainable Framework with Rich Firing Dynamics for Enhanced Temporal Processing
Authors: Alex Fulleda-Garcia, Saray Soldado-Magraner, Josep Maria Margarit Date: 2026-05-12 Paper ID: openalex:2605.11835
Summary
The paper presents multi-timescale conductance spiking networks, a novel SNN framework that achieves both rich dynamical behaviors and high activity sparsity. By parameterizing neuronal conductances, the model supports direct backpropagation without relying on surrogate gradients, addressing a major bottleneck in SNN training. Evaluation on the Mackey-Glass time-series regression task shows superior performance compared to traditional LIF and AdLIF models while maintaining lower computational overhead.
Key Contributions
- Introduces a gradient-trainable spiking network architecture using parameterized I-V curves for multi-timescale conductance control.
- Enables direct backpropagation through time without the need for surrogate gradients, improving model trainability for continuous-valued tasks.
- Outperforms baseline LIF and AdLIF networks in Mackey-Glass time-series regression while significantly reducing spike-based computational and communication activity.
Key Concepts
- multi-timescale-conductance-spiking-networks: A gradient-trainable spiking neural network framework that achieves rich firing dynamics and high sparsity by parameterizing neuronal conductance timescales.
Archivist Review
I approved the Multi-timescale Conductance Spiking Networks concept as it introduces a novel, reusable mechanism for enabling direct backpropagation in SNNs via conductance parametrization. I rejected the Mackey-Glass dataset as it is a standard mathematical benchmark rather than a primary empirical data source requiring a vault entry. No new open questions were provided or identified that meet the high standard for permanence.
Approved Concepts
- Multi-timescale Conductance Spiking Networks: This framework enables direct backpropagation through time in spiking neural networks without surrogate gradients by shaping neuronal I-V curves through multi-timescale conductances.
Rejected Candidates
- [dataset] Mackey-Glass (
mackey-glass) - low_impact: This is a standard mathematical benchmark problem for time series, not a distinct, standalone dataset entry.
Links
Metadata & Links
- url
- https://arxiv.org/abs/2605.11835
- paper_id
- 2605.11835
- paper_source
- openalex
- domain
- nlp
- tags
- nlp
- architectures
-
- datasets
-
- concept_slugs
- multi-timescale-conductance-spiking-networks
- dataset_slugs
-
- skill
- TimeSeriesSkill
- created_at
- 2026-05-15T06:16:28Z