| name | stp-stabilizes-goal-conditioned-dynamics |
| description | Short-Term Synaptic Plasticity (STP) stabilizes goal-conditioned dynamics in PFC-inspired reservoir model for multistep goal-directed action planning. Preserves action-relevant goal information under noise with 89.2% success rate vs 49.5% without STP. Activation: short-term synaptic plasticity, goal-conditioned dynamics, reservoir computing, PFC model, goal-directed planning, dynamic connectivity, facilitation-dominant STP. |
| category | neuroscience |
Context
The prefrontal cortex (PFC) maintains goal information for action planning, but how recurrent circuits preserve it in an action-usable form over behavioral timescales remains unclear. This paper demonstrates that short-term synaptic plasticity (STP) can stabilize goal information as action-usable, goal-conditioned dynamics through dynamic modulation of goal-dependent effective recurrent connectivity.
Paper: arXiv:2606.03481 (Submitted 2 Jun 2026)
Authors: Jin Nakamura, Yuichi Katori
Categories: Neurons and Cognition (q-bio.NC), Neural and Evolutionary Computing (cs.NE)
Journal: Submitted to Neural Networks (68 pages, 33 figures, 3 tables)
Core Methodology
- STP-Enhanced Reservoir Model: Incorporate STP into PFC-inspired reservoir computing model with basal-ganglia-inspired temporal-difference readout learning
- Goal-Conditioned Dynamics: Preserve goal information in action-relevant form during delay periods
- Facilitation-Dominant STP: Use STP time constants in facilitation-dominant range identified by grid search
- Effective Connectivity Analysis: Analyze goal-specific patterning of effective connectivity during delay period
- Dynamic Recurrent Modulation: History-dependent synaptic modulation stabilizes goal representations under noise
Implementation Steps
- Build Reservoir Model: Create PFC-inspired reservoir computing network with recurrent connections
- Add STP Dynamics: Implement short-term synaptic plasticity with facilitation and depression mechanisms
- Temporal-Difference Readout: Implement basal-ganglia-inspired learning for action value estimation
- Goal Representation: Encode goal identity in reservoir state during delay period
- Evaluate Noise Robustness: Test with state noise comparing models with and without STP
- Time-Resolved Decoding: Analyze when and how goal information remains decodable
- Effective Connectivity Mapping: Measure goal-specific connectivity patterns over time
Key Results
- Success Under Noise: With STP: 91.8% (no noise) → 89.2% (noise); Without STP: 75.8% → 49.5% (paired Cohen's dz=1.31)
- Goal Decodability: Goal identity highly decodable during delay even without STP
- Action-Relevance: STP preserves goal information as action-relevant goal-conditioned dynamics
- Dynamic Connectivity: Delay-period goal-specific patterning increases toward later trial phase with STP
- Time-Invariant vs Dynamic: Without STP, effective connectivity is time-invariant; with STP, it's goal-conditioned and task-state-dependent
Facilitation-Dominant Range
- Grid search identified facilitation-dominant range of STP time constants associated with high success rates
- STP state perturbation controls support online, history-dependent synaptic modulation
- Gain-matched controls argue against simple fixed recurrent-scaling explanation
Pitfalls
- STP Parameter Tuning: STP time constants require grid search for facilitation-dominant range
- Reservoir Initialization: 100 independently generated networks needed for statistical significance
- Delay Duration: STP benefits most evident during longer delay periods with noise
- Goal-Task Interaction: Goal-conditioned patterning depends on both goal and task state
- Temporal-Difference Learning: Basal-ganglia-inspired readout requires proper action-value estimation
Verification
- Compare success rates with vs without STP under noise (target: ~90% vs ~50%)
- Perform time-resolved decoding of goal information across delay period
- Analyze state-space separability between goals
- Measure action-value-difference availability at action opportunities
- Conduct STP-state perturbation and gain-matched controls
- Analyze effective connectivity patterns over time
- Grid search for facilitation-dominant STP time constants
Activation
- short-term synaptic plasticity
- goal-conditioned dynamics
- reservoir computing
- PFC model
- goal-directed planning
- dynamic connectivity
- facilitation-dominant STP
- effective connectivity
- delay period
- noise robustness
- action-relevant representation