Rethinking Spiking Neural Networks from an Ensemble Learning Perspective
Authors: Yongqi Ding, Lin Zuo, Mengmeng Jing, Pei He, Hanpu Deng
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments in 1D, 2D, and 3D scenarios demonstrate the effectiveness, generalizability, and performance advantages of our method. Our contribution can be summarized as follows: Extensive experiments on neuromorphic 1D speech/2D object recognition, and 3D point cloud classification tasks confirm the effectiveness, versatility, and performance advantages of our method. With only 4 timesteps, we achieved 83.20% accuracy on the challenging CIFAR10-DVS dataset. |
| Researcher Affiliation | Academia | Yongqi Ding, Lin Zuo , Mengmeng Jing, Pei He, Hanpu Deng School of Information and Software Engineering University of Electronic Science and Technology of China |
| Pseudocode | Yes | Algorithm 1 Temporally adjacent subnetwork guidance for SNNs Algorithm 2 Py Torch-style code for randomly dropping guidance losses |
| Open Source Code | No | The paper provides "Pytorch-style code for the drop function drop( ), which randomly drops guidance losses in the Algorithm 2" within the appendix, but does not explicitly state that the full methodology's code is open-source, nor does it provide a specific repository link for the work described in this paper. It mentions using "officially released training code directly" for Spiking Resformer, and using "code released by (Ren et al., 2023)" for point cloud classification, but these are references to other works' code, not their own. |
| Open Datasets | Yes | We perform experiments on the neuromorphic datasets CIFAR10-DVS (Li et al., 2017), DVS-Gesture (Amir et al., 2017), and N-Caltech101 (Orchard et al., 2015), as well as the static image datasets CIFAR10, CIFAR100, and the 3D point cloud datasets Model Net10 (Wu et al., 2015) and Model Net40 (Wu et al., 2015). The Spiking Heidelberg Digits (SHD) (Cramer et al., 2022) dataset contains 1000 spoken digits in 20 categories (from 0 to 9 in English and German) for the speech recognition task. To validate the scalability of our method, we performed experiments on Tiny-Image Net, Image Net Hard (Taesiri et al., 2023), and Image Net. |
| Dataset Splits | Yes | There are 10 classes of samples in CIFAR10-DVS, and we divide each class of samples into training and test sets in the ratio of 9:1 to evaluate the model performance... DVS-Gesture (Amir et al., 2017) dataset contains event samples for 11 gestures, of which 1176 are used for training and 288 are used for testing... There are 8709 samples in N-Caltech101, and we divide the training set and the test set at a ratio of 9:1. |
| Hardware Specification | Yes | Our experiments are based on the PyTorch package, using the Nvidia RTX 4090 GPU. |
| Software Dependencies | No | The paper mentions using "PyTorch package" and "Spiking Jelly (Fang et al., 2023a) framework" but does not specify their version numbers. |
| Experiment Setup | Yes | We trained the model for 100 epochs using a stochastic gradient descent optimizer with an initial learning rate of 0.1 and a tenfold decrease every 30 epochs. We trained the VGG-9 and Res Net-18 models without using any data augmentation techniques, and the weight decay value was 1e-3. The batch size during training is 64. The firing threshold ϑ and membrane potential time constant τ of spiking neurons were 1.0 and 2.0, respectively. ... TKL is the temperature hyperparameter set to 2. ... In this paper, P is set to 0.5. ... where γ is the coefficient for controlling the guidance loss, which is set to 1.0 by default. ... where a is the hyperparameter that controls the shape of the rectangular function and is set to 1.0. |