Parameter-Efficient Fine-Tuning for Spiking Point Cloud Models

2026-07-31

AI Summary

This work addresses the high computational and memory costs of full-parameter fine-tuning in existing spiking point cloud models, which also suffer from information loss due to binary spikes discarding subthreshold task-relevant signals. To overcome these limitations, we propose SpikePEFT, the first parameter-efficient fine-tuning framework tailored for spiking neural networks (SNNs) processing point clouds. By freezing pretrained synaptic weights and optimizing only approximately 5% of neuron-intrinsic parameters, SpikePEFT significantly reduces adaptation overhead while preserving the energy efficiency of SNNs. The approach introduces Intrinsic Dynamics Tuning (IDT) to modulate membrane potential decay and firing thresholds, complemented by Silent-State Disambiguated Adaptation (SSDA) to extract task-relevant features from silent neurons. Our method achieves state-of-the-art accuracy of 92.4% on ModelNet40 and 85.6% on ScanObjectNN (PB_T50_RS), demonstrating both efficacy and efficiency in downstream adaptation.

Abstract

Spiking Neural Networks (SNNs) offer energy-efficient solutions for point cloud analysis on resource-constrained devices through event-driven computation. However, existing pre-trained spiking point cloud models rely on full fine-tuning for downstream task adaptation, incurring substantial parameter and storage overhead. Furthermore, binary spike propagation suppresses task-relevant sub-threshold information. To address these issues, we propose SpikePEFT, the first parameter-efficient fine-tuning framework for spiking point cloud models. Specifically, Intrinsic Dynamics Tuning (IDT) adaptively modulates membrane decay and firing thresholds, enabling efficient neuron-intrinsic adaptation while keeping the pre-trained synaptic transformations frozen. Moreover, Silent-State Disambiguation Adaptation (SSDA) recovers task-relevant information from informative silent states, thereby providing richer evidence for downstream adaptation. Extensive experiments across multiple benchmarks demonstrate the effectiveness and efficiency of SpikePEFT. In particular, our method achieves 92.4% accuracy on ModelNet40 and 85.6\% on the most challenging classification split ScanObjectNN(PB\_T50\_RS) while updating only about 5% of the trainable parameters and preserving the energy efficiency of SNNs. This work provides a promising step toward parameter-efficient adaptation of neuromorphic vision models.

Problem

• Spiking Neural Networks
• Parameter-Efficient Fine-Tuning
• Point Cloud Analysis
• Task Adaptation
• Sub-threshold Information

Innovation

• Parameter-Efficient Fine-Tuning
• Spiking Neural Networks
• Point Cloud Analysis
• Intrinsic Dynamics Tuning
• Silent-State Disambiguation