ShiftLIF: Efficient Multi-Level Spiking Neurons with Power-of-Two Quantization

Technical Summary: ShiftLIF

Problem Statement
Spiking Neural Networks (SNNs) are increasingly recognized for edge sensing applications due to their event-driven computation and inherent temporal filtering capabilities. However, standard Leaky Integrate-and-Fire (LIF) neurons suffer from a severe representational bottleneck: they compute rich analog membrane potentials but encode them into binary spikes (0 or 1). This extreme quantization discards fine-grained amplitude information accumulated during temporal integration, limiting the network's ability to capture the complexity of continuous sensory signals.

Existing attempts to resolve this via multi-level spiking neurons face two primary limitations:

1. Hardware Inefficiency: Many multi-level approaches rely on uniformly spaced spike levels with arbitrary integer amplitudes. This reintroduces costly synaptic multiplications (MAC operations), undermining the energy efficiency that defines SNNs.
2. Distribution Mismatch: Uniform quantization grids allocate equal resolution across the entire dynamic range. This poorly matches the intrinsic distribution of membrane potentials in SNNs, which are typically heavily concentrated near small values with a long tail toward larger values. Consequently, uniform grids waste resolution on rare large activations while under-representing the densely clustered small activations where most information resides.

Methodology: ShiftLIF
The authors propose ShiftLIF, a novel multi-level spiking neuron designed to enhance representational capacity while preserving hardware-friendly, multiplier-free computation.

• Logarithmic Power-of-Two Quantization: Instead of uniform spacing, ShiftLIF maps continuous membrane potentials to a discrete set of spike levels defined by powers of two: $\{0, 2^{-K}, 2^{-(K-1)}, \dots, 2^{-1}, 1\}$. This logarithmic grid provides finer resolution near zero, aligning naturally with the heavy-tailed distribution of membrane potentials observed in trained SNNs.
• Multiplier-Free Synaptic Computation: By constraining spike amplitudes to powers of two and representing synaptic weights as integers, the synaptic interaction $W \cdot S$ is implemented as a bitwise right-shift operation ($W \gg k$) followed by integer accumulation. This eliminates general multiplication, maintaining the core energy efficiency of standard binary SNNs.
• Proportional Soft Reset: To preserve residual sub-threshold charge not encoded in the current spike, ShiftLIF employs a proportional soft-reset mechanism where the membrane potential is decremented by the magnitude of the emitted multi-level spike ($V_t \leftarrow V_t - S_t \cdot V_{th}$), rather than a fixed threshold subtraction.
• Training Mechanism: The non-differentiable shift-quantization operator is trained using a Straight-Through Estimator (STE) that approximates the derivative over the full bounded activation range $[0, 1]$. Additionally, an L1-style spike rate regularizer is applied to control firing activity without suppressing sparse responses.

Key Contributions

1. Identification of the Accuracy-Efficiency Dilemma: The paper highlights that existing multi-level SNNs often trade hardware efficiency for representational capacity due to uniform quantization and multiplication requirements.
2. ShiftLIF Architecture: The proposal of a neuron that achieves distribution-aware representation via power-of-two quantization while strictly preserving MAC-free, shift-based synaptic computation.
3. Comprehensive Cross-Modal Benchmark: The establishment of a benchmark across 10 datasets spanning wireless, acoustic, motion, and visual sensing modalities to evaluate the proposed method against state-of-the-art baselines.

Experimental Results
The authors evaluated ShiftLIF on 10 datasets (including ARIL, UrbanSound8K, UCI-HAR, and CIFAR10-DVS) using various backbones (LeNet, ResNet, Transformers) and compared it against nine other neuron models (including standard LIF, PLIF, GLIF, and INT-LIF).

• Accuracy: ShiftLIF achieved the highest average accuracy (89.34%) across all datasets, outperforming the strongest competing baselines (CLIF at 88.74% and I-LIF at 88.63%) and significantly surpassing standard binary LIF (82.00%).
• Continuous vs. Event-Based Sensing: The performance gains were most pronounced on continuous sensing tasks (wireless, acoustic, motion), where ShiftLIF consistently ranked first or tied for first. On neuromorphic vision tasks, it remained competitive.
• Energy Efficiency: Despite using multi-level spikes, ShiftLIF maintained synaptic energy consumption close to that of standard binary LIF models. In several wireless sensing benchmarks, it achieved higher accuracy with lower energy consumption compared to both binary LIF and the multi-level INT-LIF.
• Quantization Analysis: Theoretical analysis and empirical measurements confirmed that the logarithmic grid yields lower quantization error and better bit utilization (entropy) for membrane potentials concentrated near zero compared to uniform grids.

Significance and Claims
The paper claims that ShiftLIF offers a favorable trade-off between accuracy and efficiency for cross-modal edge sensing. By aligning the quantization scheme with the natural distribution of membrane potentials and leveraging bit-shift operations, ShiftLIF improves spike-level expressiveness without sacrificing the hardware-friendly nature of SNNs. The authors conclude that the performance gains stem not merely from increasing the number of spike levels, but from utilizing a quantization scheme specifically matched to SNN membrane dynamics, making it a simple and effective improvement for edge sensing applications.