Spike-based intelligence on neuromorphic chips has attracted substantial research interest due to its energy efficiency from event-driven computing on spikes. Spiking neural networks (SNNs), which make abstractions of biological neuronal networks, are the foundation for spike-based intelligence, and while they have been demonstrated in simple learning tasks, such as classification, the design of high-performance learning algorithms for SNNs, especially spiking deep learning, remains challenging. One of the main barriers is the lack of full-stack platforms that can support different hardware architectures, various datasets, biologically plausible learning rules, automatic differentiation, as well as simulation efficiency for spike-based operation. In a recent work, Guoqi Li, Yonghong Tian and colleagues addressed this gap by introducing a platform that provides convenient SNN construction, fast SNN simulation, and deployment of SNNs on different neuromorphic chips.
The proposed framework — called SpikingJelly — consists of various modules with the purpose of achieving a balance between ease of use, flexible extensibility, and high performance. In addition to providing the essential elements for building SNN structures, such as neurons and synapses, SpikingJelly also includes practical functions for training, simulating, analyzing, converting, quantizing, and deploying SNNs. Notably, the platform provides support for code generation: semi-automatically generated CUDA (Compute Unified Device Architecture) kernels can be used to accelerate the SNN simulation process. Finally, classic and large-scale network structures, such as Spiking ResNet, are made available to users, thus enabling model reuse. The authors demonstrated the flexibility and efficiency of SpikingJelly by applying it to three distinct real-world scenarios. It is also worth mentioning that the framework can be used to deploy pretrained SNNs on different neuromorphic chips and architectures. All in all, SpikingJelly has the potential to serve as an ecosystem for the coordinated development of spiking deep learning.
This is a preview of subscription content, access via your institution