Training Guide

SpikeSEG is trained with unsupervised layer-wise STDP and Winner-Take-All competition. No labelled data or backpropagation is needed.

Quick Start

python scripts/train.py --config configs/config.yaml

CLI Arguments

Flag	Default	Description
--config, -c	--	Path to YAML config file
--paper	--	Use a paper preset (igarss2023, kheradpisheh2018)
--output, -o	./runs	Output directory
--name, -n	auto	Experiment name
--resume, -r	--	Checkpoint to resume from
--epochs	50	Maximum epochs per layer
--n-classes	1	Number of output classes
--device	cuda	Device (cuda, cpu, cuda:0)
--seed	42	Random seed
--log-level	INFO	DEBUG, INFO, WARNING, ERROR
--no-tensorboard	false	Disable TensorBoard logging

Layer-Wise Training Flow

flowchart TD
    START[Load Config + Data] --> L1
    L1["Conv1: Fixed DoG/Gabor filters (no STDP)"] --> L2
    L2["Conv2: STDP + WTA until convergence"] --> L3
    L3["Conv3: STDP until convergence"] --> SAVE
    SAVE[Save Final Checkpoint]

1. Conv1 uses fixed Difference-of-Gaussians (DoG) or Gabor filters. Set train_conv1: false.
2. Conv2 is trained with STDP and WTA competition. After each sample, the WTA winner's receptive field weights are updated via STDP. Adaptive thresholds prevent dead neurons.
3. Conv3 is trained similarly. This layer learns class-level features.

Convergence

Training for each layer stops when:

• The convergence metric $C_l < 0.01$ (weights polarised to 0 or 1), or
• A target ratio of converged neurons is reached (default 90%), or
• Maximum epochs are exhausted.

Outputs

Path	Content
checkpoints/checkpoint_epoch{N}_{layer}.pt	Per-epoch checkpoint
checkpoints/checkpoint_best.pt	Best checkpoint (lowest $C_l$)
checkpoints/checkpoint_latest.pt	Last checkpoint
metrics_final.json	Per-layer statistics (wins, weight stats, convergence)
logs/	Timestamped log files
events/	TensorBoard logs (if enabled)

Programmatic Training

from spikeseg.models import SpikeSEG
from spikeseg.learning import STDPLearner, STDPConfig, WTAInhibition, WTAConfig, WTAMode

model = SpikeSEG.from_paper("igarss2023", n_classes=1)
config = STDPConfig.from_paper("igarss2023")
learner = STDPLearner(config)
wta = WTAInhibition(
    config=WTAConfig(mode=WTAMode.BOTH, local_radius=2, enable_homeostasis=True),
    n_channels=36, spatial_shape=(32, 32),
)

for epoch in range(50):
    for events, labels in dataset:
        model.reset_state()
        for t in range(10):
            output = model.encode(events[t])
            filtered, winner_mask = wta(
                output.layer_spikes['conv2'],
                output.layer_membranes['conv2'],
            )
            if winner_mask.any():
                learner.update_weights_for_winner(
                    model.encoder.conv2.weight,
                    pre_spike_times=output.layer_spike_times['conv1'],
                    post_spike_time=...,
                    winner_y=..., winner_x=...,
                    kernel_size=5,
                )
        if learner.has_converged(model.encoder.conv2.weight):
            break