Automatic Complementary Separation Pruning Toward Lightweight CNNs

TL;DR

Automatic Complementary Separation Pruning (ACSP) tackles the challenge of pruning CNNs without manual pruning volumes by combining structured pruning with activation-based pruning. It constructs a per-layer graph space that encodes component separability across all class pairs and selects a diverse, complementary subset via k-Medoids clustering, MSS scoring, and knee-point detection, enhanced by weight-aware selection. The approach is evaluated across CIFAR-10/100 and ImageNet-1K on multiple architectures, demonstrating competitive accuracy with substantial FLOPs reductions (speed-ups up to ~2.25×) and fully automated operation. This work advances practical pruning for real-world deployment by eliminating manual pruning-volume tuning and delivering hardware-friendly, scalable model compression.

Abstract

In this paper, we present Automatic Complementary Separation Pruning (ACSP), a novel and fully automated pruning method for convolutional neural networks. ACSP integrates the strengths of both structured pruning and activation-based pruning, enabling the efficient removal of entire components such as neurons and channels while leveraging activations to identify and retain the most relevant components. Our approach is designed specifically for supervised learning tasks, where we construct a graph space that encodes the separation capabilities of each component with respect to all class pairs. By employing complementary selection principles and utilizing a clustering algorithm, ACSP ensures that the selected components maintain diverse and complementary separation capabilities, reducing redundancy and maintaining high network performance. The method automatically determines the optimal subset of components in each layer, utilizing a knee-finding algorithm to select the minimal subset that preserves performance without requiring user-defined pruning volumes. Extensive experiments on multiple architectures, including VGG-16, ResNet-50, and MobileNet-V2, across datasets like CIFAR-10, CIFAR-100, and ImageNet-1K, demonstrate that ACSP achieves competitive accuracy compared to other methods while significantly reducing computational costs. This fully automated approach not only enhances scalability but also makes ACSP especially practical for real-world deployment by eliminating the need for manually defining the pruning volume.

Figures (3)

• Figure 1: Building the separation matrix for a single layer, which defines the graph space: the upper part for a linear layer, and the lower for a convolutional layer. [I] A set of samples $X$, each sample $x_i$ with a label (color). [II] Perform a forward pass to obtain activations for each sample. [III] Calculate a JM value for each component and class pair, forming a separability vector. [IV] Create the matrix representation, with each row representing a component's separability, forming the graph space.
• Figure 2: The effect of polynomial degree on the VGG-16 model on CIFAR-10 dataset. Higher polynomial degrees lead to fewer remaining FLOPs (%) but with a greater loss in accuracy.
• Figure 3: A 2-D representation of a ResNet-56 linear layer's component space, where each point represents a component, colored by its weight. The space has 7 clusters, with medoids shown as triangles and the highest-weight components in each cluster shown as rhombuses.