ShiftLIF: Efficient Multi-Level Spiking Neurons with Power-of-Two Quantization DOI: XXXXXXX.XXXXXXX Conference: Make sure to enter the correct conference title from your rights confirmation email; June 03–05, 2018; Woodstock, NY ISBN: 978-1-4503-XXXX-X/2018/06 CCS: Computing methodologies Machine learning CCS: Computing methodologies Neural networks

Kaiwen Tang, Di Yu, Jiaqi Zheng, Changze Lv, Qianhui Liu, Zhanglu Yan, Weng-Fai Wong · 2026

Read ShiftLIF: Efficient Multi-Level Spiking Neurons with Power-of-Two Quantization DOI: XXXXXXX.XXXXXXX Conference: Make sure to enter the correct conference title from your rights confirmation email; June 03–05, 2018; Woodstock, NY ISBN: 978-1-4503-XXXX-X/2018/06 CCS: Computing methodologies Machine learning CCS: Computing methodologies Neural networks by Kaiwen Tang, Di Yu, Jiaqi Zheng, Changze Lv, Qianhui Liu, Zhanglu Yan, Weng-Fai Wong (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-01866.txt.

Introduction

Tang and six colleagues design a brain-inspired neuron that speaks in more than binary spikes. Spiking neural networks suit low-power edge devices because they compute only when events arrive, but standard leaky integrate-and-fire neurons send just yes-or-no spikes, which limits what they can express. Earlier multi-level neurons widened the vocabulary, yet often used uniform steps that mismatch how membrane voltages actually distribute, or added costly multiplications. ShiftLIF maps membrane voltages onto logarithmically spaced powers of two.

That spacing gives fine detail where voltages cluster near small values, and lets synapses compute with cheap bit shifts instead of multiplications. Across ten datasets covering wireless, sound, movement, and vision sensing, ShiftLIF matched or beat existing multi-level neurons while keeping energy use close to plain binary spiking. The design offers edge sensing a better trade between accuracy and power, without giving up hardware-friendly computation.

Catalog record

CREATORKaiwen Tang, Di Yu, Jiaqi Zheng, Changze Lv, Qianhui Liu, Zhanglu Yan, Weng-Fai Wong
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE48 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.01866

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

ShiftLIF: Efficient Multi-Level Spiking Neurons with Power-of-Two Quantization DOI: XXXXXXX.XXXXXXX Conference: Make sure to enter the correct conference title from your rights confirmation email; June 03–05, 2018; Woodstock, NY ISBN: 978-1-4503-XXXX-X/2018/06 CCS: Computing methodologies Machine learning CCS: Computing methodologies Neural networks Kaiwen Tang, Di Yu, Jiaqi Zheng, Changze Lv, Qianhui Liu, Zhanglu Yan, Weng-Fai Wong Abstract . Spiking neural networks (SNNs) are promising for edge sensing due to their event-driven computation and temporal filtering capability. However, standard leaky integrate-and-fire (LIF) neurons communicate only through binary spikes, which severely limit representational capacity. Existing multi-level spiking neurons improve information transmission, but often rely on uniform quantization that mismatches membrane-potential distributions or introduces costly synaptic multiplications. In this paper, we propose ShiftLIF , a multi-level spiking neuron that maps membrane potentials to a logarithmically spaced power-of-two spike set. This design provides finer representation in the small-amplitude regime, where membrane potentials are densely concentrated, while enabling multiplier-free synaptic computation through bit-shift and accumulation operations. As a result, ShiftLIF improves spike-level expressiveness without sacrificing the hardware-friendly nature of standard SNN computation. We evaluate ShiftLIF on 10 datasets spanning wireless, acoustic, motion, and visual sensing tasks. Results show that ShiftLIF consistently matches or exceeds the accuracy of existing multi-level spiking neurons while maintaining synaptic energy consumption close to standard binary LIF. These results indicate that ShiftLIF provides a favorable accuracy-efficiency trade-off for cross-modal edge sensing. 1. Introduction The rapid expansion of edge computing has driven an increasing demand for sensing applications across wireless, acoustic, and motion domains ( 20 ; 34 ; 44 ) . However, processing these continuous streams on edge devices poses a challenge to energy consumption ( 29 ) . Spiking neural networks (SNN) have emerged as a highly promising paradigm to address this ( 51 ; 53 ; 39 ) . The most widely adopted mechanism in these networks is the leaky integrate-and-fire (LIF) neuron ( 41 ) . Beyond providing event-driven energy efficiency, the leaky integration dynamics intrinsically act as a temporal low-pass filter ( 11 ; 4 ) . This enables the neuron to smooth out high-frequency environmental noise and effectively capture slowly varying temporal patterns in real-world physical signals. Despite this advantage, standard LIF neurons suffer from a severe representational bottleneck ( 47 ) . Fundamentally, spiking neurons compute rich analog membrane potentials but encode them with only one bit. Because neurons communicate strictly through binary spikes, this extreme quantization discards much of the amplitude information preserved by temporal integration ( 16 ; 27 ) . As a result, the expressive capacity of the network is significantly constrained, limiting its ability to capture the full complexity of continuous sensory signals. Recent efforts to address this limitation generally follow two directions. The first line of work enhances neuron expressiveness by modifying membrane dynamics, such as PLIF ( 9 ) , GLIF ( 50 ) , CLIF ( 19 ) , and RPLIF ( 24 ) , which introduce learnable time constants or additional state variables. While these approaches enrich temporal modeling, they increase computational overhead and still constrain neuron outputs to binary spikes, leaving the quantization bottleneck unresolved. The second line of work increases representational bandwidth by introducing multi-level spikes, such as LM-HT SNN ( 17 ) and INT-LIF ( 27 ) . However, these methods typically rely on uniformly spaced spike levels, which introduce two limitations: multi-bit spike amplitudes require synaptic multiplications, increasing hardware cost, and uniform quantization poorly matches the intrinsic distribution of membrane potentials, which are often concentrated near small values with a long tail toward larger ones. Consequently, uniform grids allocate excessive resolution to rare large activations while under-representing densely clustered small activations. To address these challenges, we propose ShiftLIF, a novel spiking neuron that both improves representational capacity and hardware efficiency. ShiftLIF maps the continuous membrane potential to a logarithmically spaced set of spikes defined by powers of two, \{0,2^{-K},2^{-(K-1)},\dots,2^{-1},1\} . This design provides two key advantages. Algorithmically, the power-of-two grid is naturally denser near zero, which better matches the heavy-tailed distribution of membrane potentials and captures fine-grained variations in small activations that uniform quantization often misses ( 14 ; 31 ) . From a hardware perspective, because spike amplitudes are restricted to powers of two and synaptic weights are represented as integers, spike–weight interactions can be implemented using simple bit-shift operations followed by integer accumulation ( 40 ; 25 ) . Consequently, ShiftLIF enables fully multiplier-free synaptic computation, preserving the core energy efficiency advantage of spiking neural networks. Moreover, since the bit-shift operation effectively reduces the bit-width of the shifted weights, the subsequent accumulation involves smaller operands, which can further reduce switching activity and energy consumption compared to standard accumulation ( 36 ; 5 ) . To evaluate ShiftLIF, we conduct extensive experiments across 11 diverse datasets encompassing wireless, acoustic, motion, and visual sensing modalities. Empirical results demonstrate that ShiftLIF consistently achieves competitive or superior classification accuracy compared to state-of-the-art multi-level spiking neurons. Furthermore, it accomplishes this while consuming lower synaptic energy comparable to that of standard binary LIF models. The main contributions of this paper are: (1) We identify the fundamental accuracy–efficiency dilemma in existing multi-level SNNs, where uniform quantization increases computational overhead and poorly matches the natural distribution of membrane potentials. (2) We propose ShiftLIF, a novel spiking neuron that generates multi-level spikes constrained to powers of two, enabling distribution-aware representation while preserving multiplier-free synaptic computation. (3) We establish a comprehensive cross-modal benchmark across wireless, acoustic, motion, and visual sensing tasks, demonstrating that ShiftLIF achieves state-of-the-art accuracy while significantly improving energy efficiency over other LIF variants. Figure 1 . Comparison of standard LIF, existing solutions, and ShiftLIF. In (c), the colored dashed lines mark the thresholds of different power-of-two output levels, and the top inset shows a small-value-dominant membrane distribution with quantization points denser near zero. At the bottom, W\gg k denotes a bitwise right shift by k bits, highlighting that ShiftLIF replaces multiplication with shift-and-accumulate computation. 2. Related Work 2.1. SNNs for Edge Sensing Data Spiking Neural Networks have been increasingly explored for edge sensing applications, including acoustic recognition, wearable motion analysis, event-based vision, and wireless sensing, due to their event-driven sparsity and native temporal processing capability ( 6 ; 32 ; 3 ; 26 ) . Existing sensing-oriented SNN studies have primarily emphasized architectural adaptation, temporal feature extraction, encoding strategies, and surrogate gradient for direct training ( 30 ; 13 ; 28 ) . In comparison, the spike representation mechanism itself, especially the design of the neuron output alphabet beyond binary communication, remains relatively underexplored in SNNs for sensing applications design. 2.2. LIF Neurons and Their Variants To overcome the representational limitations of standard LIF dynamics, numerous variants have been proposed, broadly falling into two categories. The first category enriches internal membrane dynamics while strictly retaining binary spike communication. For instance, PLIF ( 9 ) introduces learnable membrane time constants, GLIF ( 50 ) incorporates complex gating mechanisms, and CLIF ( 19 ) adds complementary paths for better gradient flow. While these variants significantly improve temporal modeling and optimization behavior, they ultimately compress the rich, continuous membrane state into a 1-bit binary spike, leaving the inter-layer communication bottleneck unresolved. The second category directly expands the spike alphabet to mitigate this

Read the complete text