SA-MLP: A Low-Power Multiplication-Free Deep Network for 3D Point Cloud Classification in Resource-Constrained Environments

TL;DR

The paper tackles the problem of energy-efficient, real-time point cloud classification on resource-constrained sensors by replacing multiplications with lightweight shift and additive operations. It introduces Shift-MLP, Add-MLP, and the hybrid SA-MLP, with SA-MLP interleaving shift and adder layers under a tailored optimization regimen that treats each layer type differently. Empirical results on ModelNet40 show SA-MLP achieving 93.9% accuracy, surpassing the baseline Mul-MLP and other multiplication-free variants, while demonstrating robust performance across varying point densities and offering substantial computational advantages. The work highlights practical implications for edge devices in autonomous systems and robotics, and points to hardware and framework optimizations as a key direction for broader adoption of multiplication-free point-cloud models.

Abstract

Point cloud classification plays a crucial role in the processing and analysis of data from 3D sensors such as LiDAR, which are commonly used in applications like autonomous vehicles, robotics, and environmental monitoring. However, traditional neural networks, which rely heavily on multiplication operations, often face challenges in terms of high computational costs and energy consumption. This study presents a novel family of efficient MLP-based architectures designed to improve the computational efficiency of point cloud classification tasks in sensor systems. The baseline model, Mul-MLP, utilizes conventional multiplication operations, while Add-MLP and Shift-MLP replace multiplications with addition and shift operations, respectively. These replacements leverage more sensor-friendly operations that can significantly reduce computational overhead, making them particularly suitable for resource-constrained sensor platforms. To further enhance performance, we propose SA-MLP, a hybrid architecture that alternates between shift and adder layers, preserving the network depth while optimizing computational efficiency. Unlike previous approaches such as ShiftAddNet, which increase the layer count and limit representational capacity by freezing shift weights, SA-MLP fully exploits the complementary advantages of shift and adder layers by employing distinct learning rates and optimizers. Experimental results show that Add-MLP and Shift-MLP achieve competitive performance compared to Mul-MLP, while SA-MLP surpasses the baseline, delivering results comparable to state-of-the-art MLP models in terms of both classification accuracy and computational efficiency. This work offers a promising, energy-efficient solution for sensor-driven applications requiring real-time point cloud classification, particularly in environments with limited computational resources.

Figures (6)

• Figure 1: Architecture of the Multiplication-Based MLP (Mul-MLP) Model. This diagram illustrates the structure of the Mul-MLP model, where all layers utilize traditional multiplication operations for feature extraction and classification.
• Figure 2: Architecture of the Shift-Based MLP (Shift-MLP) Model. This diagram shows the structure of the Shift-MLP model, where all layers employ bitwise shift operations to achieve computational efficiency.
• Figure 3: Architecture of the Addition-Based MLP (Add-MLP) Model. This diagram presents the structure of the Add-MLP model, which replaces traditional multiplication operations with adder layers throughout the network.
• Figure 4: Architecture of the ShiftAdd-MLP (SA-MLP) Model. This diagram depicts the structure of the SA-MLP model, integrating both adder and shift layers to leverage the complementary strengths of each operation.
• Figure 5: t-SNE visualization of the coded features from the test set for each model: (a) Mul-MLP (baseline), (b) Shift-MLP, (c) Add-MLP, and (d) SA-MLP. The visualizations highlight the feature distributions and clustering behavior of the models, with SA-MLP demonstrating tighter clustering and fewer outliers.