Self-Supervised Mixer-to-Spike Learning: A Hybrid Spiking-Neural Network for Evidence-Pres

96.25% accuracy on MNIST and 92.27% accuracy on the LC25000 lung subset were achieved by the Self-Supervised Mixer-to-Spike Learning (SMSL) framework. The dataset, authored by Ashif Al Nayem Zeesan and hosted on Harvard Dataverse, was last updated on 2026-07-21. It contains images for evaluating a hybrid spiking-neural network designed for cross-domain classification.

Use Cases

• Benchmarking spiking neural network performance based on the reported accuracy and F1-scores for MNIST and LC25000.
• Training hybrid SNN-ANN models based on the described modular framework combining convolutional encoders and fusion probes.
• Studying interpretable, spike-based inference based on the framework's use of a spiking label layer for population-level competition.
• Evaluating cross-domain applicability of vision models based on tests on both handwritten digits and histopathology images.

Strengths

• Performance metrics are provided, including 96.25% accuracy on MNIST and 92.27% accuracy on LC25000.
• The framework demonstrated a 28.50 percentage point improvement over embedding-only classification on MNIST.
• The spiking readout preserved performance, reducing accuracy by only 0.32 and 0.46 percentage points compared to the non-spiking probe.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• The dataset's specific file formats and size are not detailed in the provided metadata.

Provenance

Source

Harvard Dataverse

Collection Method

Likely contains benchmark image datasets (MNIST, LC25000 lung subset) used for model evaluation.

Freshness

Last updated 2026-07-21 16:25:39; freshness should be verified.