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..
HetSyn: Versatile Timescale Integration in Spiking Neural Networks via Heterogeneous Synapses
Authors: Zhichao Deng, Zhikun Liu, Junxue Wang, Shengqian Chen, Xiang Wei, Qiang Yu
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate that Het Syn LIF not only improves the performance of SNNs across a variety of tasks including pattern generation, delayed match-to-sample, speech recognition, and visual recognition but also exhibits strong robustness to noise, enhanced working memory performance, efficiency under limited neuron resources, and generalization across timescales. In addition, analysis of the learned synaptic time constants reveals trends consistent with empirical observations in biological synapses. These findings underscore the significance of synaptic heterogeneity in enabling efficient neural computation, offering new insights into brain-inspired temporal modeling. |
| Researcher Affiliation | Academia | Zhichao Deng1,2,3 , Zhikun Liu1,2,3 Junxue Wang1,2,3 , Shengqian Chen1,2,3, Xiang Wei1,2,3, Qiang Yu1,2 1School of Artificial Intelligence, CCA Lab, Tianjin University, Tianjin, China 2College of Intelligence and Computing, Tianjin University, Tianjin, China 3School of Computer Science and Technology, Tianjin University, Tianjin, China EMAIL |
| Pseudocode | No | The paper describes the dynamics of the LIF neuron and Het Syn LIF using mathematical equations (Eq. 1-9) but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks, nor does it present the steps in a structured, code-like format. |
| Open Source Code | Yes | Code available at: https://github.com/dzcgood/Het Syn. |
| Open Datasets | Yes | We compare our proposed Het Syn LIF model with other existing works on the SHD, S-MNIST [47], Ti Digits [48], and Ti46-Alpha [49] datasets, and report the results in Table 1. |
| Dataset Splits | Yes | The SHD dataset comprises 10,420 high-quality audio samples of spoken digits (0-9) in English and German. It includes 12 speakers 6 female and 6 male aged between 21 and 56, with each speaker contributing approximately 40 utterances per digit for each language. The dataset is divided into training and testing sets, containing 8,156 and 2,264 samples, respectively. Before feeding into neural networks, we first align all audio recordings to a fixed duration of 1000 ms by trimming or zero-padding, and then sample the resulting spike trains using a 4 ms time bin, yielding a 250 700 input matrix per recording (250 timesteps 700 input channels). --- For S-MNIST, where each 28 28 image is reshaped into a sequence of 784 inputs. At each timestep, a single pixel is presented to the model in a row-wise scan from top-left to bottom-right. The MNIST dataset consists of grayscale images of handwritten digits (0 9), with 60,000 training and 10,000 test samples. We follow the original train/test split in our evaluation. --- We use the adults subset of the Ti Digits dataset, which comprises isolated utterances of the 11 English digit classes ( zero nine and oh ), with standard training and testing splits of 2,464 and 2,486 speech samples. |
| Hardware Specification | Yes | All experiments are conducted using NVIDIA RTX 4090 and Tesla V100PCIE-16GB GPUs. |
| Software Dependencies | No | The paper mentions specific optimizers (Adam, AdamW) and learning rate schedulers (Step LR, cosine annealing) but does not provide specific version numbers for any software libraries, frameworks (e.g., PyTorch, TensorFlow), or programming languages used. |
| Experiment Setup | Yes | Further experimental and training details are provided in the Appendix. --- We adopt a 100-100-1 network architecture for all models, comprising 100 input channels, 100 hidden neurons, and a single leaky-integrate readout neuron. The synaptic time constant τs is fixed at 20 ms for Hom Neu LIF and Hom Neu ALIF, and initialized by sampling from N(20, 5) ms for Het Neu LIF and Het Syn LIF. The reset time constant τJ is uniformly set to 20 ms across all models. For Hom Neu ALIF, the adaptation time constant and adaptation strength are set to τa = 500 ms and a = 0.01, respectively. All decay factors are constrained to the range [0, 1] during training. For the surrogate function, we use a triangular-shaped derivative defined as zt V t = γ max 0, 1 V t ϑ ζ ϑ , we set γ = 1 and ζ = 1. Training is performed using the Adam optimizer with a Step LR learning rate scheduler, where the learning rate is initialized at 1e-3 and decayed by a factor of 0.8 every 100 iterations. All models are trained for 1000 iterations with a batch size of 1. --- Table 2: Experiment settings and hyperparameter configurations for different datasets (includes learning rate, dropout rate, epochs, batch size, warmup ratio, optimizer, weight decay, architecture, hidden neuron number, ϑ, t, τJ, initialization of τs). |