Evolutionary Neural Architecture Search for 3D Point Cloud Analysis

TL;DR

This paper tackles the challenge of applying neural architecture search to 3D point-cloud analysis by introducing SHSADE-PIDS, a framework that encodes discrete architectures into a continuous space and searches it with an adaptive differential evolution algorithm. By integrating Success-History-based Self-adaptive Differential Evolution (SHSADE) with Joint Point Interaction Dimension Search (PIDS), the method jointly optimizes point interactions and feature dimensions while using a dense-sparse predictor and bi-objective evaluation to balance accuracy and efficiency. Empirical results on SemanticKITTI and ModelNet40 demonstrate state-of-the-art efficiency-accuracy trade-offs, achieving 64.51% mean IoU with only 0.55M parameters and 4.5 GMACs for segmentation, and 93.4% accuracy with 1.31M parameters for classification, outperforming larger hand-designed and NAS models. The findings highlight the potential of hybrid evolutionary algorithms for discrete-to-continuous neural architecture optimization in unstructured data domains, offering practical benefits for fast, deployable point-cloud models.

Abstract

Neural architecture search (NAS) automates neural network design by using optimization algorithms to navigate architecture spaces, reducing the burden of manual architecture design. While NAS has achieved success, applying it to emerging domains, such as analyzing unstructured 3D point clouds, remains underexplored due to the data lying in non-Euclidean spaces, unlike images. This paper presents Success-History-based Self-adaptive Differential Evolution with a Joint Point Interaction Dimension Search (SHSADE-PIDS), an evolutionary NAS framework that encodes discrete deep neural network architectures to continuous spaces and performs searches in the continuous spaces for efficient point cloud neural architectures. Comprehensive experiments on challenging 3D segmentation and classification benchmarks demonstrate SHSADE-PIDS's capabilities. It discovered highly efficient architectures with higher accuracy, significantly advancing prior NAS techniques. For segmentation on SemanticKITTI, SHSADE-PIDS attained 64.51% mean IoU using only 0.55M parameters and 4.5GMACs, reducing overhead by over 22-26X versus other top methods. For ModelNet40 classification, it achieved 93.4% accuracy with just 1.31M parameters, surpassing larger models. SHSADE-PIDS provided valuable insights into bridging evolutionary algorithms with neural architecture optimization, particularly for emerging frontiers like point cloud learning.

Figures (4)

• Figure 1: Proposed framework of Neural Architecture Search for 3D Point Cloud Learning Tasks via SHSADE with the encoding of discrete architectures in continuous space, Joint Point Interaction Dimension Search (PIDS), and dense-sparse predictor.
• Figure 2: Pipeline of the proposed SHSADE-PIDS for 3D point cloud, which mainly comprises the design of model evaluation, search strategy, and search space.
• Figure 3: The evolutionary process of the proposed SHSADE and the regularized EA used in PIDS (NAS) zhang2023pids.
• Figure 4: One frame from Semantic segmentation result of sequence 08 in SemanticKITTI.