Traditional Spiking Neural Networks (SNNs) are often celebrated for their biological plausibility and energy efficiency. By communicating via discrete, binary "spikes" (0 or 1), they avoid the power-hungry continuous activations of standard Artificial Neural Networks (ANNs). However, this binary nature creates a fundamental information bottleneck: a single spike carries very little data. To represent a precise value, a neuron must fire repeatedly over many simulation time steps, leading to high latency and vanishing efficiency gains.
A growing body of research is resolving this by moving beyond binary pulses toward multi-level spiking. By allowing neurons to emit multi-bit signals or multiple spikes within a single time step, these researchers are bridging the performance gap between SNNs and ANNs while achieving ultra-low latency.
The Information Bottleneck
In a standard SNN, the intensity of a signal is typically encoded through "rate coding"—the number of spikes over time. If you need to represent a value with high precision, you need many time steps. This is the "discretization error" problem: when the simulation window is too short, the network loses the nuance of the underlying data.
To combat this, Yuetong Fang and colleagues in Adaptive Calibration propose allowing neurons to emit multiple spikes per timestep. This mechanism allows the network to approximate high-precision continuous values even within very short simulation windows, effectively "calibrating" the SNN to match its ANN counterpart's accuracy.
Multi-Level Thresholds and Multi-Bit Spikes
The most direct way to increase information density is to change what a "spike" actually is. Instead of a single threshold that triggers a binary 1, researchers are implementing hierarchical or graded firing mechanisms.
Zecheng Hao and Zhaofei Yu introduced the LM-HT SNN (Learnable Multi-hierarchical Threshold), which uses equidistant multi-level thresholds. In this model, a neuron's output isn't just "on" or "off"; it's a multi-bit signal determined by how many threshold levels the membrane potential has crossed. This allows the network to achieve ImageNet-level accuracy in as few as two time steps.
Similarly, Yongjun Xiao and Xianlong Tian propose a Multi-Bit Mechanism that replaces binary spikes with fixed-point unsigned binary representations. This includes "integer bits" for multiple threshold crossings and "fractional bits" that allow neurons to transmit partial signals even if they haven't reached a full threshold. This approach preserves the "addition-only" computational advantage of SNNs while drastically reducing information loss.
Temporal Compression and Bursting
If a single time step can carry more information, we can "compress" the temporal dimension. This is often inspired by biological "bursting," where a neuron fires a rapid sequence of spikes in response to strong stimuli.
Guobin Shen and Yi Zeng developed the LIFB (Leaky Integrate and Fire or Burst) neuron (Exploiting High Performance SNNs). By adding a high-intensity "burst" state to the standard spike/no-spike options, they enable the network to function effectively in a single simulation step. This concept of temporal compression is also seen in the work of Xiaotian Song, whose One-step Spiking Transformer uses dual-threshold neurons to simulate multi-level rate coding, collapsing multiple timesteps into one without losing the representational power of the spikes.
The Bit-Budget Equivalence
Why does this work so well? Guobin Shen and Dongcheng Zhao provide a unifying mathematical perspective in Are Conventional SNNs Really Efficient?. They argue that an SNN running for $T$ time steps is mathematically analogous to a $T$-bit quantized ANN.
Under this "Bit Budget" framework, the choice becomes clear: you can either have high temporal resolution (many time steps of 1-bit spikes) or high instantaneous precision (one time step of multi-bit spikes). The research suggests that for static data like images, increasing the bit-width of individual spikes while reducing temporal steps is often more energy-efficient and accurate than traditional high-latency SNN approaches.
Applications in Edge AI
This shift toward multi-level spiking is particularly relevant for hardware-constrained environments. Dengyu Wu and Bipin Rajendran demonstrated this in Neuromorphic Wireless Split Computing, where multi-level spikes were used to balance accuracy and communication overhead in gesture recognition tasks. By increasing the information density per pulse, they achieved the low latency required for real-time edge AI.
By treating the "spike" not as a rigid binary constraint but as a flexible unit of information, these researchers are transforming SNNs from slow biological emulations into high-speed, high-precision competitors to traditional deep learning.