Conventional deep networks pass continuous real numbers between layers. Spiking neural networks instead mimic biological neurons: information is carried by the timing of discrete, binary spikes. Because a neuron only does work when it spikes, computation is event-driven and extremely energy-efficient — the basis of neuromorphic computing.
Working principle
The workhorse model is the Leaky Integrate-and-Fire (LIF) neuron. Its membrane potential integrates incoming weighted spikes and leaks toward rest over time. When the potential crosses a threshold the neuron emits a spike and resets. Information is encoded as spike rate or precise timing; learning can use biologically-inspired STDP or surrogate-gradient backpropagation through time.
| Property | ANN | SNN |
|---|---|---|
| Signal | Continuous values | Binary spikes over time |
| Time | Static forward pass | Temporal dynamics |
| Energy | Dense MAC operations | Event-driven, sparse |
| Hardware | GPU/TPU | Neuromorphic (Loihi, etc.) |
| Training | Backprop | STDP / surrogate gradients |
Trade-offOn neuromorphic chips, SNNs can cut inference energy by one to two orders of magnitude for sparse, event-based workloads — but training accuracy still lags ANNs, the field's main open problem.
Applications
- Always-on keyword spotting and gesture recognition on tiny power budgets
- Event-camera (DVS) vision for high-speed, low-latency robotics
- Edge sensor fusion where milliwatts matter
References & further reading
- Maass, “Networks of spiking neurons: the third generation of neural network models,” Neural Networks, 1997.
- Davies et al., “Loihi: A Neuromorphic Manycore Processor with On-Chip Learning,” IEEE Micro, 2018.
- Neftci et al., “Surrogate Gradient Learning in Spiking Neural Networks,” IEEE SPM, 2019.