Neuromorphic Decision Transformer (SNN-DT)

Local Plasticity, Phase-Coding, and Dendritic Routing for Low-Power Sequence Control

Abstract

This repository contains the official PyTorch implementation of the Spiking Decision Transformer (SNN-DT), as presented in the paper "Spiking Decision Transformers: Local Plasticity, Phase-Coding, and Dendritic Routing for Low-Power Sequence Control" (Pandey & Biswas, 2025).

The SNN-DT architecture bridges the gap between the sequential modeling capabilities of Transformers and the energy efficiency of Spiking Neural Networks (SNNs). By embedding Leaky Integrate-and-Fire (LIF) neurons within the self-attention mechanism and utilizing accurate STDP-inspired local plasticity, this model achieves state-of-the-art performance on continuous control tasks while reducing energy consumption by orders of magnitude compared to traditional ANN-based Decision Transformers.

*(Note: Visualizations available in the `visualizations/` directory)*

Key Features

• Neuromorphic Efficiency: Replaces standard activation functions with temporal spike-based logic, significantly reducing computational overhead suitable for edge deployment.
• Phase-Coded Positional Encoding: A biologically plausible method for encoding sequence order using spike timing phases.
• Dendritic Routing: Efficient information routing mechanism mimicking biological dendritic trees.
• Three-Factor Local Plasticity: Implements STDP-like learning rules for robust weight updates without heavy backpropagation costs during inference-time adaptation.
• Standard Gym Benchmarks: Evaluated on classic control tasks: CartPole-v1, Pendulum-v1, MountainCar-v0, and Acrobot-v1.

Installation

System requirements: Linux/Windows, Python 3.8+, CUDA-enabled GPU (recommended).

# Clone the repository
git clone https://github.com/Vishal-sys-code/neuromorphic_decision_transformer.git
cd neuromorphic_decision_transformer

# Create a virtual environment (optional but recommended)
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

Usage

Training

To train the SNN-DT model on a specific environment (e.g., Pendulum-v1), use the provided training script. The training pipeline handles data generation, preprocessing, and model optimization.

# Run training for Pendulum-v1
python snn-dt/scripts/train.py --model snn_dt --env "Pendulum-v1" --save-dir "results/snn_dt_pendulum"

To run the full suite of experiments across all environments:

./run_all_experiments.sh

Evaluation

Evaluate a pre-trained checkpoint to measure return, spike counts, latency, and estimated energy consumption.

python eval_snn_dt.py \
    --env "Pendulum-v1" \
    --checkpoint_path "results/snn_dt_pendulum/best_model.pt" \
    --target_return -200 \
    --episodes 50

Key Arguments:

• --context_len: Context length ($K$) for the transformer (default: 20).
• --per_spike_energy: Estimated energy per spike in Joules (default: 4.6pJ for 45nm process).

Project Structure

neuromorphic_decision_transformer/
├── configs/            # YAML configuration files for experiments
├── snn-dt/             # Core SNN-DT source code
│   ├── src/            # Model definitions and extensive util libraries
│   └── scripts/        # Training and utility scripts
├── demos/              # Demonstration notebooks and videos
├── eval_snn_dt.py      # Standalone evaluation script
├── requirements.txt    # Python dependencies
└── run_all_experiments.sh # Batch experiment runner

Citation

If you use this code or find our work helpful, please cite our paper:

@article{pandey2025spiking,
  title={Spiking Decision Transformers: Local Plasticity, Phase-Coding, and Dendritic Routing for Low-Power Sequence Control},
  author={Pandey, Vishal and Biswas, Debasmita},
  journal={arXiv preprint arXiv:2508.21505},
  year={2025}
}

License

This project is licensed under the MIT License - see the LICENSE file for details.