Dual Threshold Spiking Neurons for Signed Value Representation

15 concepts8 papers1 proposal

The Sign Bottleneck in Neuromorphic Computing

Traditional Spiking Neural Networks (SNNs) operate on a simple, biologically inspired principle: neurons integrate incoming signals over time into a membrane potential, and once this potential crosses a specified positive threshold, the neuron fires a discrete, positive binary spike. While this event-driven mechanism is highly energy-efficient, it introduces a fundamental architectural limitation: it is inherently unidirectional. Traditional spikes can only represent non-negative quantities.

This creates a severe mismatch when attempting to interface SNNs with modern deep learning paradigms. Artificial Neural Networks (ANNs) rely heavily on continuous, signed values. Activation functions like LeakyReLU and tanh yield negative outputs; reinforcement learning agents must track negative action-values; and optimization algorithms depend on bidirectional gradients to update weights.

As noted by Lusen Zhao et al. in their work on spiking Transformers, spikes are traditionally discrete, positive events. Without a mechanism to distinguish between positive and negative magnitudes, converting pre-trained models that rely on signed activations and weights leads to severe information loss. Similarly, Marios Karamanis et al. point out that translating linear transfer functions and representing negative quantities presents a significant challenge for standard spiking architectures, which struggle to natively encode these continuous or signed values efficiently.

The Mechanism: Dual-Threshold Integrate-and-Fire

To bridge this gap, a architectural evolution has emerged: the dual-threshold or Signed Integrate-and-Fire (SIF) neuron model. Instead of monitoring a single positive threshold ($+V_{th}$), the neuron maintains two distinct firing boundaries: a positive threshold ($+V_{th}$) and a negative threshold ($-V_{th}$).

          Membrane Potential (V)
                   |
  +V_th -----------|--------  <- Fires Positive Spike (+1)
                   |   /\
                   |  /  \
  -----------------|--------  <- Baseline (0)
                   | /    \
  -V_th -----------|/-------  <- Fires Negative Spike (-1)
                   |

The intuition behind this model is straightforward: the membrane potential acts as a bidirectional accumulator.

  • When the neuron receives positive inputs or experiences an upward shift in features, the membrane potential rises. If it crosses $+V_{th}$, the neuron emits a positive spike ($+1$) and its potential is reset downward.
  • Conversely, when the neuron receives negative inputs or experiences a decline in activity, the membrane potential drops. If it falls below $-V_{th}$, the neuron emits a negative spike ($-1$) and its potential is reset upward.

By allowing the network to emit both positive and negative spikes, the neuron can natively communicate signed magnitudes. For example, Adarsh Kumar Kosta et al. introduce a SIF neuron model in Fusion-FlowNet: Energy-Efficient Optical Flow Estimation using Sensor Fusion and Deep Fused Spiking-Analog Network Architectures to generate signed outputs. The authors demonstrate that this capability mitigates the "dead neuron" problem found in traditional Integrate-and-Fire neurons, enabling far more effective learning in event-based processing tasks like optical flow estimation.

Why It Works: Preventing Activation Collapse

The primary reason the dual-threshold mechanism succeeds is that it preserves the functional integrity of continuous activation functions during ANN-to-SNN conversion.

When converting a pre-trained ANN to an SNN, standard conversion methods typically clamp negative activations to zero because traditional spiking neurons cannot process them. Qiang Yu et al. demonstrate in Constructing Accurate and Efficient Deep Spiking Neural Networks With Double-Threshold and Augmented Schemes that this zero-clamping causes a catastrophic drop or total collapse in model performance for complex classification tasks. Discarding the negative activation spectrum strips the network of its representational capacity. By incorporating a negative firing boundary alongside the standard positive one, the spiking neurons can accurately map ANNs that utilize activation functions with negative ranges, achieving nearly lossless accuracy.

Similarly, Xu Wang et al. address this challenge in A Fast and Accurate ANN-SNN Conversion Algorithm with Negative Spikes. They show that accurately converting ANNs with Leaky ReLU activations necessitates an SNN neuron model capable of generating negative spikes. This allows for the direct representation of negative activation values, bypassing the accuracy degradation that plagues traditional non-negative spiking constraints.

Unlocking Advanced Architectures and Real-World Dynamics

By moving beyond purely positive pulses, dual-threshold neurons allow neuromorphic hardware to support complex, non-monotonic functions and dynamic real-world features across various domains:

  • Spiking Transformers: In TTFSFormer: A TTFS-based Lossless Conversion of Spiking Transformer, Lusen Zhao et al. leverage a generalized neuron structure to achieve the first lossless conversion of pre-trained Transformers into SNNs, preserving the signed weights and activations critical to self-attention mechanisms while slashing energy consumption.
  • Scientific Machine Learning: To solve differential equations on neuromorphic hardware, networks must support non-monotonic activation functions like tanh. In Spiking Physics-Informed Neural Networks on Loihi 2, Bradley H. Theilman et al. configure spiking neurons on Intel's Loihi 2 chip with dual firing thresholds. This approach allows the network to emit graded spikes that represent signed magnitudes, successfully bridging the gap between continuous physical activations and discrete event-driven hardware.
  • Dynamic Graph Learning: Real-world network dynamics involve both surges and declines in activity. Dong Chen et al. implement a dual-threshold mechanism in SiGNN: A Spike-induced Graph Neural Network for Dynamic Graph Representation Learning. This bidirectional sensitivity allows the system to capture both positive and negative fluctuations in graph features, ensuring that decreases in activity or negative feature shifts are not ignored.

Enabling Purely Spike-Based Learning

Beyond inference and conversion, signed spikes solve a fundamental bottleneck in how SNNs are trained. While SNNs run efficiently on neuromorphic hardware, they are typically trained on power-hungry GPUs using continuous floating-point math. This irony stems from the fact that backpropagation requires calculating gradients, which are inherently signed quantities.

In SPIDE: A purely spike-based method for training feedback spiking neural networks, Mingqing Xiao et al. explain that standard spiking neurons typically only communicate through positive pulses, creating a challenge for representing signed values such as gradients. Without a mechanism to convey negative information through spikes, it is difficult to implement purely spike-based learning rules that require bidirectional updates. By enabling neurons to transmit negative information via signed spikes, the network can propagate gradient updates using the same event-driven fabric it uses for forward inference, opening the door to fully on-chip, energy-efficient neuromorphic training.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_signed_value_representation_in_spiking_n_3a4b927a,
  title = {signed value representation in spiking networks},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/signed-value-representation-in-spiking-networks/rdr_3a4b927a994c4bc3ad2037934ada79f7},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_signed_value_representation_in_spiking_n_3a4b927a_v1,
  title = {signed value representation in spiking networks},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/signed-value-representation-in-spiking-networks/rdr_3a4b927a994c4bc3ad2037934ada79f7/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake