Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Enhanced Self-Distillation Framework for Efficient Spiking Neural Network Training
Authors: Xiaochen Zhao, Chengting Yu, Kairong Yu, Lei Liu, Aili Wang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on CIFAR-10, CIFAR-100, CIFAR10-DVS, and Image Net demonstrate that our method reduces training complexity while achieving high-performance SNN training. |
| Researcher Affiliation | Academia | 1 ZJU-UIUC Institute, Zhejiang University 2 College of Information Science and Electronic Engineering, Zhejiang University EMAIL |
| Pseudocode | Yes | Algorithm 1 An algorithm with caption Require: SNN model fsnn, timesteps T, hyper-parameter β, τfor rate-based backpropagation, Train datasets D = {xi, yi}n i=1, ANN block A = {A1, A2, ..., AL 1} Ensure: Train the SNN model based on self-distillation 1: for each batch training data Di = {xi, yi} do Spiking Forward 2: Update eligibility traces et ij 3: end for 4: for each batch training data Di = {xi, yi} do Rate Forward 5: Compute each sub-model outputs {p1, p2, ..., p L} 6: Aggregate the teacher labels yteacher (8) 7: Compute the self-distillation loss as shown in Equation(10) 8: Compute the CE loss as shown in Equation (9) 9: Compute the final loss using Equation (11) 10: Approximate backpropagation with eligibility traces et ij. 11: end for |
| Open Source Code | Yes | Our code is available at https://github.com/Intelli Chip-Lab/enhanced-self-distillation-framework-for-snn. |
| Open Datasets | Yes | The CIFAR-10 and CIFAR-100 [36] datasets consist of color images with a resolution of 32 32 pixels and cover a range of object categories. Both datasets are released under the MIT license. The Image Net-1K dataset [9] consists of 1,281,167 training images and 50,000 validation images across 1,000 distinct categories, and is available for non-commercial use. The CIFAR10-DVS dataset [40] is a neuromorphic version of CIFAR-10, consisting of 10,000 eventbased images captured by a Dynamic Vision Sensor (DVS) camera. The images have an increased spatial resolution of 128 128 pixels and are released under the CC BY 4.0 license. |
| Dataset Splits | Yes | CIFAR-10 contains 60,000 images across 10 classes, with 50,000 images for training and 10,000 for testing. The Image Net-1K dataset [9] consists of 1,281,167 training images and 50,000 validation images across 1,000 distinct categories. We split the dataset into 9,000 training images and 1,000 test images. |
| Hardware Specification | Yes | Experiments on CIFAR-10, CIFAR-100, CIFAR10-DVS, and Image Net datasets are conducted on an NVIDIA Ge Force RTX 3090 GPU. |
| Software Dependencies | No | Our experiments adopt a sigmoid-based surrogate gradient method [19] to approximate the Heaviside step function, defined ash(x, α) = 1/(1+e^(-αx)), where the parameter αis set to 4. We follow the timeapproximate backpropagation strategy from [77], and all implementations are based on the Py Torch [53] and Spiking Jelly [19] frameworks. |
| Experiment Setup | Yes | Our experiments adopt a sigmoid-based surrogate gradient method [19] to approximate the Heaviside step function, defined ash(x, α) = 1/(1+e^(-αx)), where the parameter αis set to 4. ... For all tasks, we use stochastic gradient descent (SGD) with a momentum of 0.9 [58], and apply a cosine annealing schedule [46] for learning rate adjustment. Additional hyperparameters are listed in the table 5. Table 5: Hyperparameters Settings. CIFAR-10 CIFAR-100 Image Net CIFAR10-DVS Epoch 300 300 100 300 Learning rate 0.1 0.1 0.2 0.1 Batch size 128 128 512 32 Weight decay 5e-4 5e-4 2e-5 5e-4 |