Spiking Neural Networks (SNNs) mimic the brain's energy efficiency by communicating via discrete, sparse events called spikes. In a traditional SNN, these spikes are strictly binary ($0$ or $1$). While this binary nature simplifies hardware implementation—replacing power-hungry multiply-accumulate (MAC) operations with simple additions—it introduces a fundamental trade-off between accuracy and latency.
To represent a high-precision continuous value (such as an activation value from a standard Artificial Neural Network, or ANN), a binary SNN must accumulate spikes over many time steps using rate coding. This temporal expansion causes high latency, eroding the energy advantages of neuromorphic hardware during real-time inference. When converting pre-trained ANNs to SNNs, this mismatch manifests as severe quantization error unless the simulation window is excessively long.
To break this bottleneck, a new paradigm replaces the single-threshold Leaky Integrate-and-Fire (LIF) neuron with a multi-threshold spiking neuron.
The Mechanism: Multi-Threshold Firing Dynamics
Instead of a single firing threshold $\theta$, multi-threshold neurons possess a hierarchy of multiple thresholds. When the internal membrane potential accumulates input signals, it is evaluated against these multiple levels simultaneously. If the potential crosses higher or multiple thresholds within a single time step, the neuron emits a multi-level or integer-valued spike rather than a binary pulse.
As shown by Zecheng Hao et al. in LM-HT SNN: Enhancing the Performance of SNN to ANN Counterpart through Learnable Multi-hierarchical Threshold Model, using equidistant multi-level thresholds allows neurons to output multi-bit signals within a single time step. This design closely approximates the continuous activation values of ANNs while maintaining a discrete, hardware-friendly firing structure.
Why It Works: Compressing Time via Information Density
The core intuition behind multi-threshold neurons is that they shift the burden of information representation from the temporal domain to the amplitude or multiplicity domain.
Consider trying to transmit the value 3. A binary neuron must remain active across at least three distinct time steps to communicate this intensity. In contrast, an integer-valued or multi-threshold neuron can fire a single pulse of magnitude 3 (or three co-occurring spikes) in exactly one time step.
This mechanism drastically increases the information density per spike event. Changqing Xu et al. demonstrated in Direct Training via Backpropagation for Ultra-Low-Latency Spiking Neural Networks with Multi-Threshold that this expanded encoding capacity directly resolves the high-latency bottleneck, enabling successful direct training via backpropagation with ultra-low latencies of just 1 to 2 time steps. Similarly, Chengting Yu et al. in MAP-SNN: Mapping spike activities with multiplicity, adaptability, and plasticity into bio-plausible spiking neural networks showed that integer-valued spike encoding preserves signal intensity across varying simulation granularities, mitigating the discretization errors that typically plague large simulation time steps.
Bridging the Gap in ANN-to-SNN Conversion
The benefits of multi-threshold dynamics are particularly evident when converting complex, pre-trained ANNs into SNNs without retraining. Traditional conversion methods suffer from severe performance degradation because binary spikes struggle to capture the wide, continuous dynamic ranges of advanced architectures like Transformers or Diffusion models.
By introducing multi-threshold neurons, the authors of Trajectory-Aware Spiking DiTs Conversion via Membrane Potential Error-Feedback enhanced the expressive capacity of spiking neurons, allowing them to accurately mirror the continuous activation trajectories of Diffusion Transformers (DiTs). In the context of vision transformers, Zihan Huang et al. in Towards High-performance Spiking Transformers from ANN to SNN Conversion utilized multi-threshold neurons to process richer information within a single step, minimizing the accuracy-latency trade-off.
Furthermore, this multi-level representation minimizes the quantization error inherent in mapping continuous values to discrete events. As detailed by Yuetong FANG et al. in Adaptive Calibration: A Unified Conversion Framework of Spiking Neural Networks, enabling neurons to emit multiple spikes per timestep allows networks to achieve high accuracy within exceptionally short simulation windows. To optimize this further, Zihan Huang et al. introduced a framework in Towards High-performance Spiking Transformers from ANN to SNN Conversion that leverages both positive and negative threshold levels to maximize representational capacity, achieving near-lossless conversion in as few as four time steps.
Expanding to Complex Domains and Hardware Efficiency
Beyond classification, multi-threshold neurons unlock the deployment of SNNs for high-fidelity spatial and edge computing tasks:
- Image Processing: Hebei Li et al. developed Deep multi-threshold spiking-UNet for image processing, proving that multi-threshold neurons maintain the detailed information flow required for dense pixel-level image processing within short time windows.
- Edge Computing: For edge AI and communication-constrained environments, Dengyu Wu et al. in Neuromorphic Wireless Split Computing with Multi-Level Spikes demonstrated that graded spike quantization balances accuracy against communication overhead, optimizing low-latency edge tasks.
Crucially, this added expressiveness does not destroy the underlying efficiency of neuromorphic hardware. In Differential Coding for Training-Free ANN-to-SNN Conversion, Zihan Huang et al. combined multi-threshold dynamics with a differential coding scheme to transmit changes in information, drastically lowering energy consumption. Finally, Xuerui Qiu et al. demonstrated in Efficient 3D Recognition with Event-driven Spike Sparse Convolution that integer-based LIF neurons enable multi-level spike outputs that replace standard power-hungry multiplications with addition-only operations during inference, achieving massive energy reductions across diverse 3D vision tasks.
Key Foundational Work
Towards High-performance Spiking Transformers from ANN to SNN Conversion (2024) — contributed 3 concept elements to this direction.
LM-HT SNN: Enhancing the Performance of SNN to ANN Counterpart through Learnable Multi-hierarchical Threshold Model (2024) — contributed 2 concept elements to this direction.
Deep multi-threshold spiking-UNet for image processing (2024) — contributed 2 concept elements to this direction.