Research map / Probabilistic and scientific modeling / Dynamical systems
Discovering plasticity rules that organize and maintain neural circuits
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.
Meta-learns local plasticity rules (CMA-ES over basis-function coefficients and time constants) that self-organize and maintain sparse sequential activity in a threshold-linear HVC-like network under synaptic turnover, including E→E and inhibitory plasticity.
Problem
How local, unsupervised plasticity rules can organize and maintain sequence-generating (timing) dynamics in neural circuits despite biological noise such as synaptic turnover and cell death, inspired by zebra finch HVC.
Approach
Parameterize plasticity as a weighted basis of pre/post activity filtered by decaying exponentials and synapse size, and meta-learn coefficients and time constants via CMA-ES with a loss that measures decodability of elapsed time, with synaptic turnover injected during meta-training; extend to separate rules for E→E, E→I and I→E synapses.
Key results
- Meta-learning without perturbation found a temporally asymmetric generalization of Oja's rule (three dominant terms) that reliably organizes sparse sequences; term-dropout analysis showed a sharp elbow at 3 terms.
- Meta-learned rules trained with and without synaptic turnover outperformed a rigid-bounds Hebbian sequence rule under turnover (p = 2.5e-8, Cohen's d = -0.45), but not distinct without perturbation after Bonferroni correction.
- Learning plasticity on E→I and I→E synapses improved time decoding over E→E-only rules, especially at high synaptic turnover rates; learned recurrent inhibition was homeostatic after afferent perturbations.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the meta-learned rule trained with synaptic turnover instead of the rigid-constraint sequence rule | Sequence-generating circuit · Time decoding after synaptic turnover | decoder loss: p = 2.5 x 10−8, Cohen’s d = −0.45 | Helped |
| Use the meta-learned rule trained without synaptic turnover instead of the rigid-constraint sequence rule | Sequence-generating circuit · Time decoding after synaptic turnover | decoder loss: p = 7.2 x 10−7, Cohen’s d = −0.44 | Helped |
| Use the meta-learned rule trained with synaptic turnover instead of the rigid-constraint sequence rule | Sequence-generating circuit · Time decoding without synaptic perturbation | decoder loss: p = 0.016, Cohen’s d = −0.29 | No clear effect |
| Use the meta-learned rule trained without synaptic turnover instead of the rigid-constraint sequence rule | Sequence-generating circuit · Time decoding without synaptic perturbation | decoder loss: p = 0.029, Cohen’s d = −0.30 | No clear effect |
| Add temporally asymmetric Hebbian terms ˜xᵢxⱼ and xᵢ˜xⱼ (ablation) | Meta-learned E→E plasticity · Sequence organization without perturbation | loss | Helped |
| Add the postsynaptic activity bound wᵢⱼ˜xⱼxⱼ (ablation) | Meta-learned E→E plasticity · Sequence organization without perturbation | loss | Helped |
| Add constant synaptic potentiation and a postsynaptic activity bound independent of synapse size (ablation) | Meta-learned E→E plasticity trained with turnover · Sequence organization with synaptic turnover | loss | Helped |
| Add temporally asymmetric Hebbian learning ˜xᵢxⱼ and the postsynaptic activity bound wᵢⱼ˜xⱼxⱼ (ablation) | Meta-learned E→E plasticity trained with turnover · Sequence organization with synaptic turnover | loss | Helped |
| Add E→I and I→E plasticity to E→E plasticity | Meta-learned sequence-generating circuit · Time decoding across synaptic turnover rates | decoder loss | Helped |
| Add the presynaptic activity term ˜xᵢxᵢ to E→I plasticity (ablation) | Meta-learned E→I plasticity · Sequence organization | loss | Helped |
Surprising findings
- Without perturbation, the meta-learned rule did not clearly outperform the rigid-bound baseline after multiple-comparison correction, so the advantage appears mainly under synaptic turnover.
- Learned plasticity rules were dense rather than sparse in the basis, so interpretation required perturbation-based term dropout analysis.
Limitations
- Meta-learning is computationally expensive and CMA-ES may need many epochs; different seeds yield different solutions.
- The chosen basis is dense and limits interpretability; it also excludes feedback-modulated plasticity.
- Initial connectivity strongly affects which plasticity rules succeed; only a limited number of seeds were tested.
Experimental setup
Benchmarks: time decoding loss in simulated HVC-like networks under synaptic turnover
Metrics: decoding loss (linear decoder) median loss across 100 test networks Kruskal-Wallis H test Cohen's d
Baselines: multiplicative asymmetric Hebbian rule with rigid synapse bounds (Fiete/Jun-styl meta-learned E→E-only rules no plasticity
Builds on: Confavreux et al. meta-learning of plasticity rules Oja's rule Nicola and Clopath sequence model Bengio et al. learning rule meta-learning
Closest papers by meaning
- Meta-learning families of plasticity rules in recurrent spiking networks using simulation-based inference (NeurIPS 2023 · similarity 0.92)
- A meta-learning approach to (re)discover plasticity rules that carve a desired function into a neural network (NeurIPS 2020 · similarity 0.92)
- Discovering heterogeneous synaptic plasticity rules via large-scale neural evolution (ICLR 2026 · similarity 0.90)
- Model Based Inference of Synaptic Plasticity Rules (NeurIPS 2024 · similarity 0.90)
- Learning the Plasticity: Plasticity-Driven Learning Framework in Spiking Neural Networks (NeurIPS 2025 · similarity 0.89)
- Feedback control guides credit assignment in recurrent neural networks (NeurIPS 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.