Bntt Snn Eval

bntt-snn-eval

Revisiting Batch Normalization for Training Low-latency Deep Spiking Neural Networks from Scratch — Kim et al. (2020) (arXiv:2010.01729, 2020)

What this evaluates

Evaluates the classification accuracy, inference latency, and energy efficiency of Spiking Neural Networks (SNNs) trained with Batch Normalization Through Time (BNTT) on standard image and neuromorphic datasets. It probes the model's ability to maintain high accuracy while drastically reducing time-steps and computational cost compared to ANN-SNN conversion and standard surrogate gradient methods.

Datasets

• CIFAR-10 — total 60000; splits: train (50000), test (10000)
• CIFAR-100 — total 60000; splits: train (50000), test (10000)
• Tiny-ImageNet — total 110000; splits: train (100000), val (10000)
• DVS-CIFAR10 — total 60000; splits: train (-1), test (-1)

Metrics

• Classification Accuracy (%) (primary) — range: percent
  • Percentage of correctly classified samples out of the total test or validation set size.
• Latency (time-steps) — range: other
  • Number of temporal steps the SNN processes before producing a final prediction. Reduced via early-exit when layer-wise gamma parameters drop below a threshold.
• Energy Efficiency Ratio — range: other
  • Ratio of ANN energy consumption to SNN energy consumption ($E_{ANN}/E_{method}$), calculated based on floating-point MAC operations reduced to additions for binary spike processing.

Input / output format

Input: RGB images (32×32 for CIFAR, 64×64 for Tiny-ImageNet) or discrete event streams (DVS-CIFAR10). Training inputs are augmented with random crop and horizontal flip.

Output: Predicted class label from the SNN's final layer.

Scoring recipe

correct = 0
total = 0
for images, labels in dataloader:
    predictions = model(images)
    correct += (predictions.argmax(dim=1) == labels).sum().item()
    total += labels.size(0)
accuracy = (correct / total) * 100

Common pitfalls

• Confusing training time-steps with inference latency, as the early-exit mechanism dynamically reduces steps per sample based on a gamma threshold.
• Assuming standard BN statistics apply directly to SNNs without temporal decoupling, which causes optimization failure on large datasets.
• Overlooking that DVS-CIFAR10 uses event-driven camera data rather than static frames, requiring specific neuromorphic preprocessing.

Evidence (verbatim from paper)

We evaluate our method on three static datasets (i.e., CIFAR-10, CIFAR-100, Tiny-ImageNet) and one neuromophic dataset (i.e., DVS-CIFAR10). ... Table 1: Classification Accuracy (%) on CIFAR-10, CIFAR-100, and Tiny-ImageNet.

Citation

@misc{kim2020revisitingbatchnormalization,
  title={Revisiting Batch Normalization for Training Low-latency Deep Spiking Neural Networks from Scratch},
  author={Kim et al. (2020)},
  year={2020},
  note={arXiv:2010.01729}
}

• arXiv: 2010.01729