where [] yt and [] xt are the output and input of the kernel and P and Q are the feedback and feedforward orders. Using this implementation, a neuron in the SRM model can be represented as a network of IIR filters, as shown in Figure 4. This architecture was adopted in [32]-[34] for the digital implementation of SRM neurons. C. Neural Coding and Spike Timing Neural coding is an essential part of the SNN. It refers to the way in which information is represented by discrete spikes. Neural coding is tightly coupled with the neuron model and determines the performance of the SNN and hardware implementation. Synapses PSP Membrane Potential Spikes PSP Spikes W1 W2 Wn PSP Spikes Figure 3. Spike response model. Synapse Filter S1[t] Z-1 αp α0 βq Z-1 β1 Σ F1[t] Z-1 Z-1 W1 Neuron Filter Sj[t] Z-1 Z-1 αp α0 βq Z-1 β1 Σ Fj[t] Wj Z-1 Σ O[t] V[t] U[t] Z-1 Z-1 θ -R[t] Σ -Vth Reset Voltage Output Spikes Σ LIF Neuron Exactly how the brain and sensory system encode information is not fully understood yet. Rate coding and temporal coding are two commonly used information coding in neuromorphic computing. Rate coding represents a value by the number of spikes in a unit time. It agrees with the observation that the sensory nerves' spike frequency increases as the stimulus intensity increases. Rate coding has been widely adopted. For example, most SNN models and neuromorphic hardware for image classification use rate coding, where the pixel value is represented by the number of spikes in unit time [13], [35]-[37]. However, rate coding has its limitations. First of all, it introduces latency. The firing rate cannot be determined accurately Reset Filter Figure 4. Neuron modeled by digital filters [29]. 10 IEEE CIRCUITS AND SYSTEMS MAGAZINE SECOND QUARTER 2022