[Paper Review] Recent Advances and New Frontiers in Spiking Neural Networks

Duzhen Zhang, Shuncheng Jia+1 more|arXiv (Cornell University)|Mar 12, 2022

Advanced Memory and Neural Computing4 citations

TL;DR

This survey reviews recent advances and emerging frontiers in spiking neural networks (SNNs), focusing on five core areas: neuron models, encoding methods, topology structures, neuromorphic datasets, optimization algorithms, software frameworks, and hardware platforms. It highlights SNNs' potential for energy-efficient, brain-inspired artificial intelligence through event-driven computation and neuromorphic hardware integration.

ABSTRACT

In recent years, spiking neural networks (SNNs) have received extensive attention in brain-inspired intelligence due to their rich spatially-temporal dynamics, various encoding methods, and event-driven characteristics that naturally fit the neuromorphic hardware. With the development of SNNs, brain-inspired intelligence, an emerging research field inspired by brain science achievements and aiming at artificial general intelligence, is becoming hot. This paper reviews recent advances and discusses new frontiers in SNNs from five major research topics, including essential elements (i.e., spiking neuron models, encoding methods, and topology structures), neuromorphic datasets, optimization algorithms, software, and hardware frameworks. We hope our survey can help researchers understand SNNs better and inspire new works to advance this field.

Motivation & Objective

• To provide a comprehensive review of recent progress in spiking neural networks (SNNs) across five key research domains.
• To identify and analyze emerging challenges and frontiers in SNN research, particularly in bridging biological plausibility with computational efficiency.
• To support researchers in understanding the current state of SNNs and inspire new work toward artificial general intelligence.
• To examine the role of neuromorphic hardware and software frameworks in accelerating SNN deployment and training.
• To compare and contrast offline and online neuromorphic chips in supporting SNN inference and learning.

Proposed method

• Systematically surveys recent literature across five core domains: essential elements (neuron models, encoding, topology), neuromorphic datasets, optimization algorithms, software frameworks, and hardware platforms.
• Analyzes neuron models including Hodgkin-Huxley, LIF, and Izhikevich, emphasizing their biological fidelity and computational efficiency.
• Reviews encoding methods such as rate coding, temporal coding, and rank-based coding, assessing their impact on information representation in SNNs.
• Examines optimization techniques including spike-timing-dependent plasticity (STDP) and pseudo-backpropagation (e.g., Zenke & Ganguli, 2018) for training SNNs.
• Evaluates software tools like SpikingJelly and CogSNN for SNN development and training.
• Reviews neuromorphic hardware platforms including IBM TrueNorth, Intel Loihi, and Tianjic, focusing on their architecture, scalability, and support for on-chip learning.

Experimental results

Research questions

• RQ1 What are the key advancements in spiking neuron models that balance biological realism and computational efficiency?
• RQ2 How do different spike encoding methods affect the performance and robustness of SNNs in event-based processing?
• RQ3 What are the current limitations and opportunities in optimizing SNNs using biologically inspired (e.g., STDP) and engineering-driven (e.g., pseudo-BP) algorithms?
• RQ4 How do modern software frameworks enable the development and training of SNNs, and what are their scalability and usability trade-offs?
• RQ5 What are the architectural innovations in neuromorphic hardware that enable ultra-low-power, high-parallelism SNN inference and learning?

Key findings

• The Hodgkin-Huxley model provides a detailed biophysical description of neuronal dynamics but is computationally expensive, while LIF and Izhikevich models offer efficient approximations suitable for large-scale SNNs.
• Neuromorphic datasets such as N-MNIST and DVS-CIFAR10 are essential for training and evaluating SNNs on event-based vision tasks, enabling high-speed, low-power processing.
• Pseudo-backpropagation methods like those proposed by Zenke & Ganguli (2018) enable end-to-end backpropagation in SNNs, significantly improving accuracy on benchmark tasks.
• Neuromorphic chips such as Intel Loihi (8 million neurons, 8 billion synapses) and IBM TrueNorth (1 million neurons, 256 million synapses) demonstrate ultra-low power consumption and high parallelism for real-time SNN inference.
• Online neuromorphic chips like Loihi and SpiNNaker support on-chip learning, enabling adaptive, event-driven computation, whereas offline chips like TrueNorth and Neurogrid are limited to inference-only deployment.
• Hybrid architectures such as Tianjic support both DNNs and SNNs, enabling multi-modal applications like speech recognition and obstacle avoidance in self-driving bicycles.