Research map / Architectures and efficiency / Spiking networks
Learning the Plasticity: Plasticity-Driven Learning Framework in Spiking Neural Networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PDLF, a framework that meta-learns parametric synaptic plasticity rules (SCP + PDP, expanding a spiking BCM rule) with evolution strategies in SNNs, instead of directly optimizing weights, and shows gains in working memory, multi-task RL, and robustness to damage.
Problem
Directly optimized synaptic weights in spiking neural networks yield static connections that limit adaptability, memory, multi-task learning and generalization; hand-designed local plasticity rules are rigid.
Approach
Learn the parameters of a parametric plasticity rule (Synaptic Cooperation Plasticity plus Presynaptic-Dependent Plasticity, built from a spiking BCM expansion with learnable A, B, C, D coefficients) via evolution strategies, so that synaptic weights are updated by learned local rules during the agent's lifetime rather than fixed after training.
Key results
- On ant_dir, Opt_SCP+PDP reaches 6904 ± 801 reward versus 1069 ± 98 for direct weight optimization.
- On swimmer_dir, PDLF reaches 10531 ± 827 versus 31 ± 8 for direct weight optimization.
- Removing PDP causes SNN training to diverge, showing PDP is needed for stability.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PDLF (SCP+PDP) instead of directly optimizing synaptic weights | Three-layer SNN · ant_dir | reward: 1069 ± 98 -> 6904 ± 801 | Helped |
| Use PDLF (SCP+PDP) instead of directly optimizing synaptic weights | Three-layer SNN · swimmer_dir | reward: 31 ± 8 -> 10531 ± 827 | Helped |
| Use PDLF (SCP+PDP) instead of directly optimizing synaptic weights | Three-layer SNN · halfcheetah_vel | reward: -1598 ± 324 -> -549 ± 95 | Helped |
| Use PDLF (SCP+PDP) instead of directly optimizing synaptic weights | Three-layer SNN · hopper_vel | reward: 729 ± 116 -> 869 ± 38 | Helped |
| Use PDLF (SCP+PDP) instead of directly optimizing synaptic weights | Three-layer SNN · fetch | reward: 15 ± 0.6 -> 51 ± 3 | Helped |
| Use PDLF (SCP+PDP) instead of directly optimizing synaptic weights | Three-layer SNN · ur5e | reward: 7 ± 5 -> 86 ± 5 | Helped |
| Add SCP to PDP (ablation) | PDLF · ant_dir | reward: 3284 ± 570 -> 6904 ± 801 | Helped |
| Add SCP to PDP (ablation) | PDLF · swimmer_dir | reward: 7831 ± 1479 -> 10531 ± 827 | Helped |
| Add SCP to PDP (ablation) | PDLF · halfcheetah_vel | reward: -870 ± 312 -> -549 ± 95 | Helped |
| Add SCP to PDP (ablation) | PDLF · hopper_vel | reward: 792 ± 50 -> 869 ± 38 | Helped |
| Add SCP to PDP (ablation) | PDLF · fetch | reward: 26 ± 26 -> 51 ± 3 | Helped |
| Add SCP to PDP (ablation) | PDLF · ur5e | reward: 58 ± 13 -> 86 ± 5 | Helped |
Surprising findings
- Removing the presynaptic-dependent plasticity term makes the SNNs diverge in multi-task RL.
- Direct weight-trained SNNs can end up with near-trivial behaviors such as maintaining approximate immobility on multi-direction tasks.
Limitations
- Results are reported on small, custom synthetic working-memory and simulated RL tasks with a three-layer SNN.
- Energy efficiency claims are qualitative and based on firing-rate visualizations rather than measured energy.
- Only five trials per configuration; statistical comparison across methods is limited.
Experimental setup
Datasets: Brax simulated environments (ant_dir, swimmer_dir, halfcheetah_vel, hopper_vel, ur5e, fetch) synthetic working memory copying task
Benchmarks: copying working memory task Brax multi-task RL (ant_dir, swimmer_dir, halfcheetah_vel, hopper_vel, fetch, ur5e)
Metrics: accuracy episodic reward mean and standard deviation over 5 trials
Baselines: Opt_Weight (direct weight optimization with ES) Opt_PDP (ablation) Opt_SCP (ablation) Chance level
Builds on: Spiking BCM rule STDP PEPG Brax
Closest papers by meaning
- Meta-learning families of plasticity rules in recurrent spiking networks using simulation-based inference (NeurIPS 2023 · similarity 0.92)
- Discovering heterogeneous synaptic plasticity rules via large-scale neural evolution (ICLR 2026 · similarity 0.92)
- Model Based Inference of Synaptic Plasticity Rules (NeurIPS 2024 · similarity 0.90)
- Discovering plasticity rules that organize and maintain neural circuits (NeurIPS 2024 · similarity 0.89)
- Neuroplastic Expansion in Deep Reinforcement Learning (ICLR 2025 · similarity 0.89)
- Autaptic Synaptic Circuit Enhances Spatio-temporal Predictive Learning of Spiking Neural Networks (ICML 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.