Research map / Probabilistic and scientific modeling / Dynamical systems
Model Based Inference of Synaptic Plasticity Rules
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 a gradient-based method that infers synaptic plasticity rules (truncated Taylor series or MLP) from neural activity or behavioral trajectories, validates on simulations, and applies it to fruit-fly reward-learning behavior to reveal a weight-decay (active forgetting) term.
Problem
Inferring the synaptic plasticity rules that govern learning in the brain from experimental neural or behavioral data, where synaptic weights and generative rules are not directly observed.
Approach
Parameterize the plasticity rule as a truncated Taylor series or MLP and fit its parameters by backpropagation through time over entire observed trajectories, so that the rule can depend on postsynaptic activity and current weights (which prior logistic-regression approaches could not handle).
Key results
- On simulated neural data, the method recovers Oja's rule coefficients (theta110 and theta021 approach 1 and -1) and maintains high accuracy with recordings from only 50% of neurons.
- On simulated behavior, the MLP and Taylor models reach weight R2 around 0.85 for the xjr rule and identify the relevant term, with percent deviance explained around 62-65%.
- On Drosophila behavior, adding a weight-decay term improves fit (Wilcoxon p = 5e-5) and the inferred weight coefficient is negative, consistent with active forgetting.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add a synaptic weight-dependent term (wij) to the fly plasticity rule (ablation) | Fly behavior model · Drosophila two-alternative choice task | percent deviance explained: p = 5×10−5 | Helped |
| Add reward expectation to absolute reward in a rule with weight dependence (ablation) | Fly behavior model · Drosophila two-alternative choice task | percent deviance explained: p = 0.067 | Helped |
| Vary trajectory length from 30 to 1920 (ablation) | Taylor series behavior model · Simulated behavior | percent deviance explained: 39.52 -> 79.66 | Helped |
| Vary trajectory length from 30 to 1920 (ablation) | Taylor series behavior model · Simulated behavior | R2 weights: 0.79 -> 0.64 | Hurt |
| Use shorter moving-average windows for expected reward (ablation) | Behavior simulation model · Simulated behavior | R2 weights | Helped |
| Increase hidden-layer size from 10 to 1000 (ablation) | Taylor series behavior model · Simulated behavior | R2 weights: 0.78 -> 0.79 | No clear effect |
| Use ground-truth rule xjyiwijr −0.05r instead of xjr (ablation) | Taylor series plasticity inference · Simulated reward-based plasticity | percent deviance explained: 61.91 -> 84.22 | Mixed |
| Use ground-truth rule xjyiwijr −0.05r instead of xjr (ablation) | Taylor series plasticity inference · Simulated reward-based plasticity | R2 weights: 0.78 -> 0.41 | Hurt |
Surprising findings
- Solely reward-based rules lacking E[R] or weight dependence yield high activity R2 but low weight R2 because unidirectional potentiation saturates the sigmoid.
- The method recovers weight trajectories well even though it is optimized only on neural activity or choice data.
Limitations
- Lacks complex temporal dependencies such as metaplasticity.
- Sloppiness: overparameterized models fail to identify a unique sparse solution even with extensive data.
- Connectome or architecture mismatch between generative and model networks may lead to incorrect rules.
Experimental setup
Datasets: simulated neural activity from Oja's rule network simulated behavior from covariance-based reward rule Drosophila two-alternative choice behavior (18 flies)
Benchmarks: simulated Oja's rule recovery simulated reward-based plasticity recovery Drosophila behavioral choice data
Metrics: R2 over weights R2 over activity percent deviance explained Wilcoxon signed-rank test test percent deviance explained
Baselines: Rajagopalan et al. 2023 logistic regression method Taylor series parameterization vs MLP parameterization
Builds on: Oja's rule Confavreux et al. 2020 meta-learning plasticity Rajagopalan et al. 2023 logistic regression rule inference covariance-based learning rule (Loewenstein & Seung 2006)
Closest papers by meaning
- A meta-learning approach to (re)discover plasticity rules that carve a desired function into a neural network (NeurIPS 2020 · similarity 0.93)
- Discovering heterogeneous synaptic plasticity rules via large-scale neural evolution (ICLR 2026 · similarity 0.93)
- Learning the Plasticity: Plasticity-Driven Learning Framework in Spiking Neural Networks (NeurIPS 2025 · similarity 0.90)
- Discovering plasticity rules that organize and maintain neural circuits (NeurIPS 2024 · similarity 0.90)
- Synaptic Weight Distributions Depend on the Geometry of Plasticity (ICLR 2024 · similarity 0.87)
- Flexible inference for animal learning rules using neural networks (NeurIPS 2025 · similarity 0.87)
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.