Skip to main content ホーム クリエイター hiyenwong ai_collection astrocyte-3body-plasticity
astrocyte-3body-plasticity Astrocyte-centric 3-body plasticity framework — how astrocytes participate in synaptic credit assignment alongside pre- and postsynaptic neurons. Proposes astrocytes as slow modulatory agents that bridge Hebbian millisecond-timescale plasticity with memory consolidation over seconds-to-minutes. Use when: modeling tripartite synapse learning rules, synaptic credit assignment, neuro-glial co-computation, biologically plausible learning in SNNs.
インストールへ移動 Skills Marketplace コミュニティが作成したAIスキルを発見・探索
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
直接コマンドでは確認用 Prompt が省略されます。実行前にソースを確認してください。
npx skills add https://github.com/hiyenwong/ai_collection --skill astrocyte-3body-plasticityコマンドは1行のまま表示されます。コピー前に横へスクロールして全体を確認してください。
ローカルで確認しますか？SkillsMP が現在取得できるファイルをダウンロードできます。
Zipをダウンロード ダウンロード中... name astrocyte-3body-plasticity version 1.0 description Astrocyte-centric 3-body plasticity framework — how astrocytes participate in synaptic credit assignment alongside pre- and postsynaptic neurons. Proposes astrocytes as slow modulatory agents that bridge Hebbian millisecond-timescale plasticity with memory consolidation over seconds-to-minutes. Use when: modeling tripartite synapse learning rules, synaptic credit assignment, neuro-glial co-computation, biologically plausible learning in SNNs.
tags ["astrocyte","plasticity","tripartite-synapse","credit-assignment","hebbian-learning","STDP","computational-neuroscience","synaptic-plasticity","glia","memory"] activation_keywords ["astrocyte","tripartite synapse","glia","glial plasticity","credit assignment neuroscience","3-body plasticity","astrocyte learning","slow plasticity"] source {"pmid":"42183627","journal":"The Neuroscientist","year":2026,"title":"The 3-Body Problem: How Astrocytes May Govern Plasticity","authors":["Watanabe Airi","Guo Connie","Sjöström P Jesper"]}
Astrocyte 3-Body Plasticity Framework
Overview
Learning and memory were long thought to be the exclusive domain of neurons. However, astrocytes — the predominant glial cell type — actively participate in synaptic plasticity and information storage. This skill encodes the 3-body problem framework for astrocyte-mediated plasticity: the computational and coding roles of astrocytes in the tripartite synapse (presynaptic neuron + postsynaptic neuron + astrocyte).
Key insight : While Hebbian / spike timing-dependent plasticity (STDP) operates on millisecond timescales between pre- and postsynaptic neurons, astrocytes integrate signals over seconds-to-minutes, potentially providing a slow credit assignment signal that bridges the temporal gap between spike events and behavioral outcomes.
Core Concepts
1. The Tripartite Synapse as 3-Body System
Body 1 (Pre) : Presynaptic neuron releases glutamate/GABABody 2 (Post) : Postsynaptic neuron integrates and firesBody 3 (Astrocyte) : Perisynaptic astrocytic processes detect spillover, integrate Ca²⁺ signals, release gliotransmitters Pre ──→ Synapse ──→ Post
↕
Astrocyte
(Ca²⁺ integrator)
2. Temporal Scales Process Timescale Agent Spike transmission 1–5 ms Pre → Post STDP window 10–100 ms Pre + Post Astrocyte Ca²⁺ rise 1–10 s Astrocyte Gliotransmitter release 10 s – min Astrocyte LTP/LTD consolidation min – hr All three
3. Credit Assignment Role Astrocytes may solve the distal credit assignment problem by:
Integrating neuromodulatory signals (dopamine, ACh) alongside synaptic activity
Providing a retrograde "eligibility trace" signal (e.g., via D-serine, ATP, glutamate)
Acting as a temporal bridge between fast spike events and slow reward/error signals
4. Computational Properties of Astrocytes
Spatial integration : Single astrocyte contacts 300,000+ synapses (in human cortex)Nonlinear calcium dynamics : Threshold-like activation creates binary/graded signalingIP3R-mediated Ca²⁺ stores : Endoplasmic reticulum acts as memory elementGliotransmitter release : Modulates NMDA-R via D-serine co-agonist site
Modeling Astrocyte-Mediated Plasticity
ChI Model (Chi Model for Calcium + IP3) class AstrocyteCaModel :
"""
Simplified De Pittà / Li-Rinzel astrocyte calcium model
"""
def __init__ (self ):
self .C = 0.1
self .h = 0.9
self .I = 0.1
self .d1 = 0.13
self .d2 = 1.049
self .d3 = 0.9434
self .d5 = 0.08234
self .v_C = 6.0
self .k_C = 0.1
self .C_er_0 = 2.0
def m_inf (self ):
"""IP3R open probability"""
return (self .I / (self .I + self .d1)) * (self .C / (self .C + self .d5))
def h_inf (self ):
return self .d2 / (self .C + self .d2)
def step (self, J_in, dt=0.001 ):
"""Advance astrocyte calcium by dt seconds"""
J_C = (0.6 * self .m_inf()**3 * self .h**3 *
(self .C_er_0 - self .C - self .C))
J_pump = self .v_C * self .C**2 / (self .k_C**2 + self .C**2 )
dC = J_in + J_C - J_pump
dh = (self .h_inf() - self .h) / (1.0 / (0.001 * (self .C + self .d2)))
self .C += dC * dt
self .h += dh * dt
return self .C
Tripartite STDP Rule class TripartiteSTDP :
"""
STDP modulated by astrocyte calcium signal.
Based on Bhatt-Bhatt-Bhatt-type rules with gliotransmitter gating.
"""
def __init__ (self, A_plus=0.01 , A_minus=0.012 , tau_plus=20e-3 , tau_minus=25e-3 ):
self .A_plus = A_plus
self .A_minus = A_minus
self .tau_plus = tau_plus
self .tau_minus = tau_minus
def dw_astrocyte_modulated (self, dt_spike, astro_ca, threshold=0.3 ):
"""
Compute synaptic weight change modulated by astrocyte Ca2+.
dt_spike: t_post - t_pre (positive = post after pre = LTP direction)
astro_ca: astrocyte calcium level (0 to ~1 μM normalized)
threshold: Ca2+ threshold for gliotransmitter release
"""
if dt_spike > 0 :
dw_stdp = self .A_plus * np.exp(-dt_spike / self .tau_plus)
else :
dw_stdp = -self .A_minus * np.exp(dt_spike / self .tau_minus)
if astro_ca > threshold:
gliotransmitter_factor = 1.0 + (astro_ca - threshold) * 2.0
else :
gliotransmitter_factor = astro_ca / threshold
return dw_stdp * gliotransmitter_factor
Key Experimental Evidence
Bhatt et al. 2020 : Astrocyte Ca²⁺ signals are necessary for LTP induction at CA1 synapsesMin & Bhatt 2012 : D-serine from astrocytes required for NMDA-R-dependent LTPPanatier et al. 2011 : Astrocytic D-serine controls synaptic NMDA receptor tonePerea et al. 2016 : Astrocytes mediate synapse-specific inhibitory plasticityBhatt et al. 2023 : Astrocyte IP3 signaling links neuromodulation to memory consolidation
Theoretical Framework
Credit Assignment Hypothesis The 3-body framework proposes that astrocytes solve the local credit assignment problem by:
Reward signal (e.g., dopamine)
↓
Astrocyte detects DA via D1/D2 receptors
↓
Ca²⁺ rise occurs IF (AND) synapse was recently active
↓
Gliotransmitter (D-serine / ATP) released
↓
Modifies NMDA conductance at the ACTIVE synapse
↓
NMDA-dependent plasticity rules (LTP/LTD) are gated
This provides a three-factor rule :
Δw ∝ pre × post × astrocyte(neuromodulator × activity_history)
Implications for AI/ML
Slow memory systems : Astrocyte-like "slow neurons" could implement credit propagation over longer time horizonsAttention-gated plasticity : Astrocytes as spatial filters for "important" synapsesMetabolic constraints : Astrocytes couple energy supply (glucose → lactate) to learning signals
Implementation in SNN Frameworks
Using SpikingJelly + Custom Astrocyte Module import torch
import torch.nn as nn
class AstrocyteModule (nn.Module):
"""
Astrocyte-like slow integrator for modulating plasticity.
"""
def __init__ (self, n_synapses, tau_ca=5.0 , dt=1e-3 ):
super ().__init__()
self .tau_ca = tau_ca
self .dt = dt
self .ca = nn.Parameter(torch.zeros(n_synapses), requires_grad=False )
def forward (self, synaptic_activity, neuromodulator_signal ):
"""
synaptic_activity: recent pre*post coincidence (0 or 1 per synapse)
neuromodulator_signal: scalar reward/DA signal
"""
dCa = (-self .ca / self .tau_ca +
synaptic_activity * neuromodulator_signal)
self .ca.data += dCa * self .dt
self .ca.data.clamp_(min =0 )
gate = torch.sigmoid((self .ca - 0.3 ) * 10 )
return gate
class TripartiteSynapseLayer (nn.Module):
def __init__ (self, n_in, n_out ):
super ().__init__()
self .linear = nn.Linear(n_in, n_out, bias=False )
self .astrocyte = AstrocyteModule(n_in * n_out)
def forward (self, x_pre, x_post, reward ):
activity = (x_pre.unsqueeze(-1 ) * x_post.unsqueeze(-2 )).reshape(-1 )
gate = self .astrocyte(activity, reward)
if self .training:
with torch.no_grad():
dw = (x_pre.T @ x_post) * gate.reshape(self .linear.weight.shape)
self .linear.weight.data += 0.01 * dw
return self .linear(x_pre)
Research Directions
Astrocyte-enabled continual learning : Using slow Ca²⁺ dynamics to prevent catastrophic forgettingSpatial credit assignment : Leveraging astrocyte "territory" (single cell covers 100,000 synapses) for local learning rulesNeuromodulator-gated plasticity : Dopamine + astrocyte interaction for reinforcement learningEnergy-aware learning : Astrocyte-mediated metabolic coupling to prioritize high-value synapses
Pitfalls
Timescale mismatch : Astrocyte Ca²⁺ kinetics (1–10 s) vs standard STDP (ms) requires careful temporal discretizationSpatial resolution : Single astrocyte integrates many synapses — model must decide if integration is local or broadcastCa²⁺ model complexity : Detailed IP3R models are stiff ODEs; use simplified 2-variable models for large networksGliotransmitter identity : D-serine vs ATP vs TNF-α have distinct mechanisms; model must specify which pathway
Related Skills
three-factor-snn-learning — three-factor learning rules overviewadaptive-spiking-neurons-asn — adaptive neuron modelsdual-timescale-memory-spiking-neuron-astrocyte — SNAN dual-timescale memoryatp-hysteresis-tripartite-synapse — ATP-based hysteresis in tripartite synapsesmeta-learning-biological-plasticity — meta-learning bioplausible rules 