Synchrony-Gated Plasticity with Dopamine Modulation for Spiking Neural Networks

Official PyTorch implementation of DA-SSDP (Dopamine-modulated Spike-Synchrony-Dependent Plasticity), proposed in

Synchrony-Gated Plasticity with Dopamine Modulation for Spiking Neural Networks
Transactions on Machine Learning Research (TMLR)

DA-SSDP is a light-weight, synchrony-based plasticity rule that runs during training only.
It plugs into deep spiking transformers such as SpikingResformer and adds a local update on top of surrogate backpropagation, without changing the forward architecture or inference cost.

The training scripts and configs follow the original SpikingResformer codebase.

---

1. Installation

numpy==1.24.4 Pillow==11.2.1 PyYAML==5.4.1 PyYAML==6.0.2 spikingjelly==0.0.0.0.15 thop==0.1.1.post2209072238 timm==1.0.15 torch==1.12.1 torchvision==0.13.1

---

2. Datasets

2.1 ImageNet-1K

Organize ImageNet in the standard way:

/path/to/imagenet
|-- train
|   |-- n01440764
|   |-- n01443537
|   `-- ...
`-- val
    |-- n01440764
    |-- n01443537
    `-- ...

3. Training recipes

The training pipeline (optimizer, schedules, augmentations) is the same as in SpikingResformer.
DA-SSDP is enabled or disabled through the YAML config.

3.1 ImageNet with DA-SSDP (SpikingResformer-S backbone)

Example: distributed training on 8 GPUs

torchrun \
    --standalone \
    --nnodes=1 \
    --nproc-per-node=8 \
    main.py \
    -- configs/main/spikingresformer_l.yaml \
    --data-path /path/to/imagenet \
    --output-dir /path/to/output_da_ssdp

Baseline SpikingResformer-L (no DA-SSDP) with the same setup:

torchrun \
    --standalone \
    --nnodes=1 \
    --nproc-per-node=8 \
    main.py \
    --configs/main/spikingresformer_l.yaml \
    --data-path /path/to/imagenet \
    --output-dir /path/to/output_baseline

---

3.2 Direct training on CIFAR-10 / CIFAR-100

Example: direct training on CIFAR-10 with a tiny backbone plus DA-SSDP

python main.py \
    --configs/direct_training/cifar10.yaml \
    --data-path /path/to/cifar100 \

Example: CIFAR-100

python main.py \
    --configs/direct_training/cifar100.yaml \
    --data-path /path/to/cifar100 \
    --output-dir /path/to/output_cifar100

---

3.3 Transfer learning from CIFAR10-DVS

Example: fine-tune an ImageNet-pretrained SpikingResformer-S + DA-SSDP on CIFAR10-DVS

python main.py \
    --configs/transfer/cifar10dvs.yaml \
    --data-path /path/to/cifar10 \
    --output-dir /path/to/output_cifar10_ft \
    --transfer /path/to/imagenet_checkpoint.pth

---