Static Training for Spiking Neural Networks

Instead of using multiple spiking neurons per node to achieve biological plausibility, the patent inverts the approach by using a single spiking neuron per node. The biological plausibility is maintained through the use of realistic neuron models (LIF, I&F, LIF with adaptation) and synaptic dynamics (STDP, rate-based learning) rather than through neuronal multiplication. This single-neuron-per-node design reduces device complexity while preserving biological realism through accurate modeling of neuronal and synaptic behavior.

The patent employs multiple spiking neuron models with varying degrees of biological realism to balance implementation simplicity and biological plausibility. The Leaky Integrate-and-Fire (LIF) model provides a simplified implementation while the Adaptive LIF model with adaptation currents and the I&F model with detailed synaptic dynamics provide enhanced biological plausibility. This parameter-based approach allows selection of appropriate model complexity for different applications, maintaining ease of implementation while improving biological realism when needed.

The patent uses static response functions during the training phase to achieve training efficiency, then applies these trained weights to dynamic spiking neuron implementations. The static training phase allows efficient computation using standard backpropagation, while the dynamic implementation phase captures temporal dynamics. This preliminary training approach maintains training efficiency while enabling accurate dynamic implementation through proper weight transfer and subsequent fine-tuning if needed.