NeuroTrain: A Unified Benchmarking Framework for Spiking Neural Network Training Algorithms
NeuroTrain, a newly developed open-source framework, seeks to unify the assessment of training algorithms for spiking neural networks (SNNs). Utilizing the snnTorch library, it features a modular and extendable design that incorporates a representative array of learning rules. A detailed survey accompanies it, offering a thorough classification of SNN training techniques, including surrogate-gradient backpropagation, local and three-factor learning rules, biologically inspired plasticity, ANN-to-SNN conversion methods, and unconventional optimization approaches. Each category is examined based on computational principles, learning signals, and locality characteristics. By providing a standardized interface for comparing various methods, NeuroTrain promotes reproducible research and addresses the existing gap in fine-grained taxonomy within the field, ultimately aiding in benchmarking and advancing SNN research.
Key facts
- NeuroTrain is an open-source framework based on snnTorch.
- It implements a representative set of SNN training algorithms.
- The survey provides a taxonomy of SNN training methods.
- Categories include surrogate-gradient, local learning, three-factor rules, plasticity, ANN-to-SNN conversion, and non-standard strategies.
- Each class is analyzed for computational principles, learning signals, and locality.
- The framework is modular and extendable.
- Aims to support reproducible research.
- Addresses the lack of a unified taxonomy in SNN training.
Entities
Institutions
- arXiv