Related Experiment Video
Updated: Dec 10, 2025
10:50
Computational Modeling of Retinal Neurons for Visual Prosthesis Research - Fundamental Approaches
Published on: June 21, 2022
2.1K
Comparing SNNs and RNNs on neuromorphic vision datasets: Similarities and differences
Summary
This study systematically compares Spiking Neural Networks (SNNs) and Recurrent Neural Networks (RNNs) on neuromorphic vision data. Results offer insights for selecting models and developing novel neural architectures for event-driven computing.
Area of Science:
- Neuromorphic computing
- Machine learning
- Artificial intelligence
Background:
- Neuromorphic data, characterized by frameless spike events, are crucial for understanding spatiotemporal information and event-driven processing.
- Spiking Neural Networks (SNNs) are event-driven models adept at neuromorphic computing, often benchmarked on such data.
- Recurrent Neural Networks (RNNs), while not event-driven, also possess capabilities for extracting spatiotemporal features, raising questions about direct comparisons.
Purpose of the Study:
- To systematically compare the performance of SNNs and RNNs (including vanilla RNNs and LSTMs) on neuromorphic vision datasets.
- To identify similarities and differences between SNNs and RNNs from modeling and learning perspectives.
- To provide insights for model selection and the development of novel neural models for neuromorphic applications.
Main Methods:
- Unified supervised learning using backpropagation through time (BPTT) for comparability.
- Standardized network structures (fully-connected/convolutional layers) and hyper-parameters.
- Modified RNN loss function inspired by SNN rate coding and tuned temporal dataset resolutions.
Main Results:
- Extensive contrast experiments were conducted on DVS-converted (N-MNIST) and DVS-captured (DVS Gesture) datasets.
- Insights were gained into recognition accuracy, feature extraction, temporal resolution effects, learning generalization, computational complexity, and parameter volume.
- Performance differences and similarities between SNNs and RNNs were elucidated under controlled conditions.
Conclusions:
- The study provides valuable insights for choosing appropriate models for specific neuromorphic workloads.
- Findings can guide the development of new neural network architectures tailored for event-driven data processing.
- A fair and systematic comparison framework was established for evaluating SNNs and RNNs on neuromorphic tasks.
