Reconstructing Spiking Neural Networks Using a Single Neuron with Autapses

A team of researchers, including Daqing Guo, has published a groundbreaking paper introducing the Time-Delayed Autapse Spiking Neural Network (TDA-SNN). This novel framework challenges the conventional architecture of Spiking Neural Networks (SNNs), which are prized for their energy efficiency and biological plausibility in neuromorphic computing. Instead of relying on dense, multi-layer networks with thousands of neurons, TDA-SNN reconstructs their computational power using just a single leaky integrate-and-fire neuron equipped with autapses—self-connections that create internal feedback loops with time delays. By reorganizing the neuron's internal temporal states over time, the system can emulate the functions of reservoir networks, multi-layer perceptrons, and even convolution-like architectures within a single, unified model.

The key innovation is a shift from spatial complexity (many neurons) to temporal complexity (one neuron doing more over time). Experiments on sequential, event-based, and image classification benchmarks showed that TDA-SNN achieves performance competitive with standard SNNs in reservoir and MLP settings. The model dramatically reduces the physical neuron count and the state memory required, while increasing the information capacity of each individual neuron. However, this compactness comes with a trade-off: in extreme single-neuron configurations, tasks require additional temporal latency to complete. This research, published on arXiv, fundamentally rethinks neural network design, demonstrating that sophisticated computation can be temporally multiplexed onto minimal hardware, paving the way for ultra-compact neuromorphic chips.